Wednesday, 20 April 2011

Cleanup of occurrence records

Lars here, like Oliver I've started here in October 2010 and have no biology background either so my first step here at GBIF was to set up the infrastructure Tim was mentioning before, but I've written about that already (at length).

To continue the series of blog posts that was started by Oliver, and in no particular order, I'll talk about what we are doing to process the incoming data - which is the task I was given after the Hadoop setup was done.

During our rollover we're processing Occurrence records. Millions of them, about 270 millions at the moment and we expect this to grow significantly over the next few months and years. It is only natural that there is bound to be bad data in there for various reasons. These might be everything from simple typos to misconfigured publishing tools and transfer errors.

The more we know about the domain and the data the more we are obviously able to fix. Any input is appreciated on how we could do better on this part of our processing.

For fields like kingdom, phylum, country name or basis of record we do a simple lookup in a dictionary to look for common mistakes and replace those with the proper versions. Other fields like class, order, family, genus and author have way too many distinct values for us to prepare a dictionary with all the possible errors and their correct forms. That is why we only apply a few safe clean up procedures here (e.g. remove blacklisted names or invalid characters).

Scientific names are additionally parsed by the NameParser in the ECAT project which does all kinds of fancy magic to try to infer a correct name. Altitudes, depths and coordinates get treatment as well by looking at common unit markers and errors we've seen in the past.

And last but not least we also try to make most out of the dates we get. As everyone who ever dealt with date strings knows this can be one of the hardest topics in an internationalized environment. In theory our input data consists of three nicely formatted fields: year, month and day. In reality though a lot of dates are just in the year field. We've got all kinds of delimiters (with "/" and "-" being among the most common ones), abbreviations ("Mar") and database export fragments ("1978.0" because it was a floating point variable in the database), missing data and more.

Additionally we obviously have to deal with different time formats. Is "01/02/02" the first of February or the second of January? In most cases we can only guess.

Having said that: We've rewritten large parts of the date handling routines and are continuing to improve them as we know that this is an important part of our data. Feedback on how we're doing here is greatly appreciated!

I'm really hoping to have a chance to compile a few statistics about our incoming data quality once we've tested all of this in production.

Monday, 18 April 2011

Reworking the Portal processing

The GBIF Data Portal has provided a gateway to discover and access the content shared through the GBIF network for some years, without major change.  As the amount of data has grown, GBIF have scaled vertically (e.g. scaling up) to maintain performance levels; this is becoming unmanageable with the current processing routines due to the amount of SQL statements issued against the database.  As GBIF content grows, the indexing infrastructure must change to scale out accordingly.

I have been monitoring and evaluating alternative technologies for some time and a few months ago GBIF initiated the redevelopment of the processing routines.  This current area of work does not increase functionality offered through the portal (that will be addressed following this infrastructural work) but rather aims to:
  • Reduce the latency between a record changing on the publisher side, and being reflected in the index
  • Reduce the amount of (wo)man-hours needed to coax through a successful processing run
  • Improve the quality assurance by inclusion of    
  • Rework all the date and time handling
  • Use dictionaries (vocabularies) for interpretation of fields such as Basis of Record
  • Integrate checklists (taxonomic, nomenclatural and thematic) shared through the GBIF ECAT Programme to improve the taxonomic services, and the backbone ("nub") taxonomy.
  • Provide a robust framework for future development
  • Allow the infrastructure to grow predictably with content and demand growth
Things have progressed significantly since my early investigations, and GBIF are developing using the following technologies:
    • Apache Hadoop: A distributed file system, and cluster processing using the Map Reduce framework
    • Sqoop: A utility to synchronize between relational databases and Hadoop 
    • Hive: A data warehouse infrastructure built on top of Hadoop, and developed and open-sourced by Facebook.  Hive gives SQL capabilities on Hadoop.  [Full table scans on GBIF occurrence records reduce from hours to minutes]
    • Oozie: An open-source workflow/coordination service to manage data processing jobs for Hadoop, developed then open-sourced by Yahoo!
    [GBIF are researching using HBase, the Hadoop database to allow an increase in the richness in the indexed content, and will be the subject of future blogs.  See the project site]

    The processing workflow looks like the following (click for full size):

    The Oozie workflow is still being developed, but the workflow definition can be found here.

    Lucene for searching names in our new common taxonomy

    Oliver here - I'm one of the new developers at GBIF, having started in October, 2010. With no previous experience in biology or biological classification you can bet it's been a steep learning curve in my time here, but at the same time it's very nice to be learning about a domain that's real, valuable and permanent, rather than yet another fleeting e-commerce, money-trading or "social media" application!

    One of the features of GBIF's Data Portal is allowing searching of primary occurrence data via a backbone taxonomy. For example let's say you're interested in snow leopards and would like to plot all current and historical occurrences of this elusive cat on a world map. Let's further say that Richard Attenborough suggested to you that the snow leopard's scientific name is "Panthera uncia". You would ask the data portal for all records about Panthera uncia and expect to see all occurrences of snow leopards. Unfortunately biologists aren't agreed on how to classify the snow leopard - some argue that it belongs in the genus Panthera, while others argue that it should belong to its own genus, Uncia, and naturally the GBIF network has records under both names. You would just like to see all of those records and never mind the details - and that's just the tip of the iceberg when it comes to building a backbone taxonomy to match the 260 million+ occurrence records in the GBIF network.

    Indeed, the backbone taxonomy (we call it our "Nub Taxonomy") in use by the current data portal has been one of the biggest sources of criticism of the GBIF data portal - it doesn't cover enough of the names in our occurrence records, and it doesn't handle the tricky stuff (as above) as well as it should. One of the reasons for that is the current backbone taxonomy was built based on the Catalogue of Life 2007, a similar vintage International Plant Names Index (IPNI), and then augmented with the classifications from any unmatched occurrence records. This has led to a classification hierarchy which is less reliable than we (and the GBIF network) would like.

    Markus Döring is the GBIF software team's taxonomy expert and he has employed a new strategy for building an improved Nub Taxonomy by building it exclusively on well-known and respected taxonomies already out there - things like the most recent Catalogue of Life, IPNI, and more, but without using the classifications as given in the occurrence data. After the Nub Taxonomy is built, the occurrence records then need to be matched to it. As the first step to integrating the new Nub Taxonomy into the data portal, my job in the last little while has been to build a searchable index of all the names in our Nub Taxonomy and a web service that can accept a scientific name (from an occurrence record) and match it to the index, while understanding the implications of homonyms and synonyms, as well as tolerating misspellings. And of course, make it fast :)

    Since what we're talking about here is string matching with a tolerance for messy input (e.g. spelling mistakes, different violations of nomenclatural rules) the place to start is Lucene. Our Nub Taxonomy has about 8 million unique names, and our 260 million occurrence records are also comprised of roughly 8 million unique names. Our use case is somewhat out of the ordinary for Lucene in that we can build the index once and after that it becomes read-only until the next update of our Nub Taxonomy (e.g. to reflect an update in the Catalog of Life), and it only takes a few minutes to build the index, so it's not all that important for it to be persistent. That means we can optimize for search speed and not worry so much about indexing performance. Lucene has just the index storage implementation for this need - RAMDirectory. For the most part this worked just fine, but no matter how hard I hit the index, I couldn't get cpu usage to 100% - the best I could do was about 80%. I found that very irksome and spent some time testing different Directory implementations, web service stacks, and everything in between. None of the other Directory implementations (all file based in some way) showed any improvements, nor did eliminating the web stack. Finally by attaching a profiler to the Tomcat instance running the webservice while running with RAMDirectory we were able to see thread blocking increasing proportional to the number of requesting threads. That led us to the Lucene source code where we found a synchronized() block that we deemed the culprit. With the cause at least found I decided not to waste time trying to fix the problem for what would be nominal gain, but instead decided to use two Tomcat installations and load balance between them with Apache. With the Tomcats running on quite powerful machines we are now seeing approximately 1000 lookups/sec (including a bunch of business logic beyond the Lucene lookup), which we think is pretty good, and sufficient for our purposes.

    This is all being used from within our Oozie orchestrated Hive/Hadoop workflow (which Lars will talk more about soon) but once we're confident that it's behaving properly and stably we will also offer this web service (or something similar) for public consumption. More importantly the new Nub Taxonomy will be available in the GBIF data portal very soon and with it we hope to have eliminated most of the problems people have found with our current backbone taxonomy.

    Friday, 15 April 2011

    The first drafts of the Data Publishing Manuals are available for feedbacks

    Since Darwin Core had been officially ratified by Biodiversity Information Standards (TDWG) in November 2009, a few tools were developed by GBIFS to leverage the standard data format, a.k.a the Darwin Core Archive, to facilitate data mobilisation. These tools include Darwin Core Archive Assistant, GBIF Spreadsheet Processor and some validators that users can use to produce standard-compliant files for data exchanging or publishing purposes. Also, IPT has upgraded to version 2 recently to fully support data publishing in metadata, occurrence data and taxonomic data using Darwin Core Archive.

    Accompanying these development efforts, a suite of document are also prepared to instruct users on, not only the usage of individual software tool, but how to make data available within the GBIF Network. For those tool options we have in the biodiversity information world, we organised these materials according to which kind of content that users want to publish, and present a document map for users to follow. So, if you go to the Informatics section of the GBIF web site, there are pages called "publishing" under "Discovery/Metadata," "Primary Data" and "Name Services." Maps are there ready to guide you through the way for publishing your data. Every node in the maps are clickable and will lead you to those individual manuals.

    The intention of using a map as a guide is to suggest a route that user can have basic understanding about data publishing before they play with software tools, so readers are not given a bunch of documents and don't know where to start, or hesitate to finish reading all of these before starting. We also try to keep each manual as compact as possible, with emphasis on steps, rather than just theories.

    In addition to users with biodiversity background, we'd like invite developers to evaluate these draft materials, too. Any comments are welcome, especially whether these manuals help in explaining the data publishing workflow to the users you serve.

    Wednesday, 13 April 2011

    Can IPT2 handle big datasets now?

    One of IPT1's most serious problems was its inability to handle large datasets. For example, a dataset with only half a million records (relatively small compared to some of the biggest in the GBIF network) caused the application to slow down to such a degree that even the most patient users were throwing their hands up in dismay.
    Anyways, I wanted to see for myself whether the IPT’s problems with large datasets have been overcome or not in the newest version: IPT2.

    Here’s what I did to run the test: First, I connected to a MySQL database and used a “select * from … limit …” query to define my source data totalling 24 million records (the same number of records as a large dataset coming from Sweden). Next, I mapped 17 columns to Darwin Core occurrence terms and once this was done I was able to start the publication of a Darwin Core Archive (DwC-A). The publication took just under 50 minutes to finish, processing approximately 500,000 records per minute. Take a look at the screenshot below that was taken after the successful publication. Important to note is that this test was run on a Tomcat server with only 256MB of memory. In fact, special care was taken during IPT2 design to ensure it could still run on older hardware that didn’t have a lot of memory. It’s worth noting that this is one of the reasons why IPT2 is not as feature rich as the IPT1 was.


    So just how does the IPT2 handle 24 million records coming from a database while running on a system with so little memory? The answer is that instead of returning all records at once, they are retrieved in small result sets only having about 1000 records each. These results sets are then streamed to file and immediately written to disk. The final DwC-A generated was 3.61GB in size, so some disk space is obviously needed too.

    Therefore in conclusion I feel that he IPT2 has successfully overcome its previous problems handling large datasets. I hope other adopters will now give it a shot themselves.

    Monday, 11 April 2011

    The GBIF Development Team

    Recently the GBIF development group have been asked to communicate more on the work being carried out in the secretariat.  To quote one message:
    "IMHO, simply making all these discussions public via a basic mailing list could help people like me ... have a better awareness of what's going on... We could add our comments / identify possible drawbacks / make some "scalability tests"... In fact I'm really eager to participate to this process" (developer in Belgium)
    To kick things off, we plan to make better use of this blog and have set a target of posting 2-3 times a week.  This is a technical blog, so the anticipated audience include developers, database administrators and those interested in following details of the GBIF software development.  We have always welcomed external contributers to this blog and invite any developers working on publishing content through the GBIF network, or developing tools that make use of content discoverable and accessible through GBIF to write posts.

    Today we are pleased to welcome Jan Legind to the team who will be working as a data administrator to help improve the frequency of the network crawling (harvesting) and the indexing processes.  Jan will be working closely with the data publishers to help improve the quality and quantity of content accessible through GBIF.

    The GBIF development group has expanded in the past 6 months, so I'll introduce the whole team working in the secretariat and contracted to GBIF:

    • Developers (in order of appearance in the team): Kyle Braak, José Cuadra, Markus Döring (contracted in Germany), Daniel Amariles & Hectór Tobón (contracted at CIAT in Colombia), Federico Méndez, Lars Francke and Oliver Meyn. 
    • Systems architect: Tim Robertson
    • Systems analyst: Andrea Hahn
    • Informatics liason: Burke (Chih-Jen) Ko
    • Systems admins: Ciprian Vizitiu & Andrei Cenja
    • Data administrator: Jan Legind

    The current focus of work at GBIF include the following major activities:
    • Developing and rolling out the Integrated Publishing Toolkit.
    • Integrating the checklist (taxonomic, nomenclatural and thematic) content into the current Data portal.
    • Developing a processing framework to automate the steps needed to apply quality control and index content for discovery through the Data portal.
      • Specifically to reducing the time taken and complexity in initiating a rollover of the content behind the index
      • Reworking all quality control (geographic, taxonomic and temporal) 
      • Automating the process
    • Initiating a redesign of the data portal user interface to provide richer discovery and integration across dataset metadata, checklists and primary biodiversity data.
    • Reducing the time between publishing content onto the network and discovery through the Data portal.  This includes providing specific support to those who are experiencing problems with large datasets in particular, and assisting in migration to the DarwinCore-Archive format.
    • Technical and user documentation of the publishing options available
    Let the blogging begin.

    [Please use #gbif in twitter hashtags]

    Wednesday, 26 January 2011

    Setting up a Hadoop cluster - Part 1: Manual Installation

    Introduction

    In the last few months I was tasked several times with setting up Hadoop clusters. Those weren't huge - two to thirteen machines - but from what I read and hear this is a common use case especially for companies just starting with Hadoop or setting up a first small test cluster. While there is a huge amount of documentation in form of official documentation, blog posts, articles and books most of it stops just where it gets interesting: Dealing with all the stuff you really have to do to set up a cluster, cleaning logs, maintaining the system, knowing what and how to tune etc.

    I'll try to describe all the hoops we had to jump through and all the steps involved to get our Hadoop cluster up and running. Probably trivial stuff for experienced Sysadmins but if you're a Developer and finding yourself in the "Devops" role all of a sudden I hope it is useful to you.

    While working at GBIF I was asked to set up a Hadoop cluster on 15 existing and 3 new machines. So the first interesting thing about this setup is that it is a heterogeneous environment: Three different configurations at the moment. This is where our first goal came from: We wanted some kind of automated configuration management. We needed to try different cluster configurations and we need to be able to shift roles around the cluster without having to do a lot of manual work on each machine. We decided to use a tool called Puppet for this task.

    While Hadoop is not currently in production at GBIF there are mid- to long-term plans to switch parts of our infrastructure to various components of the HStack. Namely MapReduce jobs with Hive and perhaps Pig (there is already strong knowledge of SQL here) and also storing of large amounts of raw data in HBase to be processed asynchronously (~500 million records until next year) and indexed in a Lucene/Solr solution possibly using something like Katta to distribute indexes. For good measure we also have fairly complex geographic calculations and map-tile rendering that could be done on Hadoop. So we have those 18 machines and no real clue how they'll be used and which services we'd need in the end.

    Environment

    As mentioned before we have three different server configurations. We've put those machines in three logical clusters c1, c2 and c3 and just counting up in those (our master for example is currently running on c1n1):
    • c1 10: Intel(R) Xeon(R) CPU X3363 @ 2.83GHz, 2x6MB (quad), 8 GB RAM, 2 x 500GB SATA 7.2K
    • c2 3: 2 x Intel(R) Xeon(R) CPU E5630 @ 2.53GHz (quad), 24 GB RAM, 6 x 250 GB SATA 5.4K
    • c3 5: Intel(R) Xeon(R) CPU X3220 @ 2.40GHz (quad), 4 GB RAM, 2 x 160 GB SATA 7.2K
    • CentOS 5.5
    • The machines are in different racks but connected to only one switch
    We realize that this is a very heterogeneous cluster configuration. We also realize that some people highly discourage use of old machines or machines with little RAM but the c1 and c3 clusters were old unused machines and this way they still serve a purpose and we've had no problems so far using this setup.


    Goal

    These were the goals we set out to achieve on our cluster and these are also all the things I'll try to describe in this or a following post:
    • Puppet for setting up the services and configuring machine state
    • CDH3 (Beta 3)
      • Hadoop HDFS + MapReduce incl. Hadoop LZO
      • Hue
      • Zookeeper
      • HBase
    • Easily distributable packages for Hadoop, Hive and Pig to be used by the employees to access the cluster from their own workstations
    • Benchmarks & Optimizations
    Be warned: This is going to be a very long post and unfortunately it is the nature of these things that some of the information is bound to be outdated pretty quickly so let me know if something has changed and I'll alter the post.

    Manual Installation

    Before we use Puppet to do everything automatically I will show how it can be done manually. I think it is important to know all the steps in case something goes wrong or you decide not to use Puppet at all. When I talk about "the server" I always mean "all servers in your cluster" except when noted otherwise. I highly recommend not skipping this part even if you want to use Puppet.

    Operating System

    For now I'll just assume a vanilla CentOS 5.5 installation. There's nothing special you need. I recommend just the bare minimum, everything else needed can be installed at a later time. A few words though about things you might want to do: Your servers probably have multiple disks. You shouldn't use any RAID or LVM on any of your slaves (i.e. DataNodes/TaskTracker). Just use a JBOD configuration. In our cluster all disks are in a simple structure:
    • /mnt/disk1
    • /mnt/disk2
    • ...
    There are also two tweaks for your slaves you can do:
    • Mount your data disks with noatime (e.g. /dev/sdc1 /mnt/disk3 ext3 defaults,noatime 1 2 which btw. implies nodiratime)
    • By default there are a certain number of blocks reserved on ext (not familiar with others) file systems (check by running tune2fs -l /dev/sdc1 and look at the Reserved block count). While this is useful on system disks so that critical processes can still write some data when the disk is full otherwise this is wasted space on our data disks. By default 5% of a HDD are reserved for this. I recommend setting this down to 1% by running: tune2fs -m 1 <device> on all your data disks (i.e. tune2fs -m 1 /dev/sdc1) which frees up quite a bit of disk space. You can also set it to 0% if you want though I went with 1% for our cluster. Keep the default setting for your system disks though!
    On your NameNode however use any means you feel necessary to secure your data. You know your requirements better than I do. Use RAID and/or LVM however you like. We don't have any special resources so our NameNode is running on one of our regular servers at the moment. We might change that in the future.

    A note on Cloudera's Package system & naming

    Cloudera provides the various components of Hadoop in different Packages but they follow a simple structure: There is one hadoop-0.20 package which contains all the jars, config files, directories, etc. needed for all the roles. And then there are packages like hadoop-0.20-namenode which are only a few kilobytes and they only contain the appropriate start- and stopscripts for the role in question.

    1. Common Requirements

    Most of the commands in this guide need to be executed as root. I've chosen the easy route here and just logged in as root. If you're operating as a non-privileged user remember to use su, sudo or any other means to ensure you have the proper rights.

    Repository As all the packages we're going to install are provided by Cloudera we need to add their repository to our cluster:
    curl http://archive.cloudera.com/redhat/cdh/cloudera-cdh3.repo > /etc/yum.repos.d/cloudera-cdh3.repo
    Java installation You have to download the JDK from Oracle's website yourself as license issues prevent it from being added to the repositories. Chose the correct system (probably Linux x64) and make sure to download the file ending in -rpm.bin (i.e. jdk-6u23-linux-x64-rpm.bin). You might have to do this from a client machine because you need a browser that works with the Oracle site. So on any one machine execute the following:
    unzip jdk-6u23-linux-x64-rpm.bin
    You should now have a bunch of .rpm files but you only need one of them: jdk-6u23-linux-amd64.rpm. Copy this file to your servers and install it as root using rpm:
    rpm -Uvh ./jdk-6u23-linux-amd64.rpm
    Time
    While not a hard requirement it makes a lot of things easier if the clocks on your servers are synchronized. I added this part at the last minute because we just realized that ntpd was disabled on three of our machines (c2) by accident and had some problems with it. It is worth taking a look at the clocks now and set up ntp properly before you start.

    DNS
    It doesn't matter if you use a DNS server or hosts files or any other means for the servers to find each other. But make sure this works! Do it now! Even if you think everything's set up correctly. Another thing that you should check is if the local hostname resolves to the public IP address. If you're using a DNS server you can use dig to test this but that doesn't take into account the /etc/hosts file so here is a simple test to see if it is correct:
    ping -c 1 `hostname`
    This should resolve to the public IP and not to 127.0.0.1.

    Firewall
    Hadoop uses a lot of ports for its internal and external communications. We've just allowed all traffic between the servers in the cluster and clients. But if you don't want to do that you can also selectively open the required ports. I try to mention them but they can all be changed in the configuration files. I might also miss some due to our config so I'd be glad if someone could point those out to me.

    Packages
    We're going to use lzo compression, the Hadoop native libraries as well as hue so there are a few common dependencies on all machines in the cluster which can be easily installed:
    rpm -Uvh http://download.fedora.redhat.com/pub/epel/5/x86_64/epel-release-5-4.noarch.rpm
    yum install -y lzo hue-plugins hadoop-0.20-native
    Directories
    We also need some directories later on so we can just create them now:
    mkdir <data disk>/hadoop
    chown root:hadoop <data disk>/hadoop
    Cloudera uses the alternatives system to manage configuration. In /etc/hadoop/conf is the currently activated configuration. Look at the contents of /etc/hadoop and you'll find all the installed configurations. At the moment there is only a conf.empty directory which we'll use as our starting point:
    cp -R /etc/hadoop/conf.empty /etc/hadoop/conf.cluster
    Now feel free to edit the configuration files in /etc/hadoop/conf.cluster but we'll go through them as well later in this post. The last step is to activate this configuration:
    /usr/sbin/alternatives --install /etc/hadoop-0.20/conf hadoop-0.20-conf /etc/hadoop-0.20/conf.cluster 50
    LZO Due to licensing issues the LZO bindings for Hadoop cannot be distributed the same way as the rest of the packages. So this - once again - involves a few manual steps. After these bindings were removed from Hadoop itself a few versions ago they moved tho the hadoop-gpl-compression project on Google Code which (as far as I know) still works but hasn't seen any development for over a year. Thankfully though Twitter's Kevin Weil and Cloudera's Todd Lipcon have picked up the project and maintained it. They regularly sync their github repositories so both should have almost the same code. I'm going to use Todd's version here as it should be better synced with CDH releases. You have to download the code from the repository, build the native libraries as well as the jar file and distribute those files on your cluster. You need to do this only on one machine which ideally should run the same OS version as the servers in your cluster. When you're finished you can just copy the result to all servers. We're using version 0.4.9 so we use this to download and build:
    rpm -Uvh http://download.fedora.redhat.com/pub/epel/5/x86_64/epel-release-5-4.noarch.rpm
    yum install -y lzo-devel
    wget --no-check-certificate https://github.com/toddlipcon/hadoop-lzo/tarball/0.4.9
    tar xvfz toddlipcon-hadoop-lzo-0.4.9-0-g0e70051.tar.gz
    wget http://www.apache.org/dist/ant/binaries/apache-ant-1.8.2-bin.tar.bz2
    tar jxvf apache-ant-1.8.2-bin.tar.gz
    cd toddlipcon-hadoop-lzo-0e70051
    JAVA_HOME=/usr/java/latest/ BUILD_REVISION="0.4.9" ../apache-ant-1.8.2/bin/ant tar
    The ant version that comes with CentOS 5.5 didn't work for me that's why I downloaded a new one. This should leave you with a hadoop-lzo-0.4.9.tar.gz file in the build directory which you can extract to get all the necessary files for your servers:
    • hadoop-lzo-0.4.9.jar needs to be copied into /usr/lib/hadoop/lib on each server
    • lib/native/Linux-amd64-64 needs to be copied into /usr/lib/hadoop/lib/native on each server
    cron & log cleaning
    We've had a problem with unintentional debug logs filling up our hard drives. The investigations that followed that incident resulted in a blog post by Lars George explaining all the log files Hadoop writes. It is a worthwhile read. Hadoop writes tons of logs in various processes and phases and you should make sure that these don't fill up your hard drives. There are two instances in the current CDH3b3 where you have to manually interfere:
    • Hadoop daemon logs
    • Job XML files on the JobTracker
    Hadoop uses a DailyRollingFileAppender which unfortunately doesn't have a maxBackupIndex setting like the RollingFileAppender. So either change the appender or manually clean up logs after a few days. We chose the second path and added a very simple cron job to run daily:
    find /var/log/hadoop/ -type f -mtime +14 -name "hadoop-hadoop-*" -delete
    This jobs deletes old log files after 14 days. We'll take care of the Job XML files in a similar way at the JobTracker.

    HDFS

    One property needs to be set for both the NameNode and the DataNodes in the file /etc/hadoop/conf/core-site.xml: fs.default.name. So just add this and replace $namenode with the IP or name of your NameNode:
    <property>
      <name>fs.default.name</name>
      <value>hdfs://$namenode:8020</value>
    </property>

    2.1. NameNode
    Installing the NameNode is straightforward:
    yum install -y hadoop-0.20-namenode
    This installs the startup scripts for the NameNode. The core package was already installed in the previous step. Now we need to change the configuration, create some directories and format the NameNode.

    In /etc/hadoop/conf/hdfs-site.xml add the dfs.name.dir property which "determines where on the local filesystem the DFS name node should store the name table(fsimage). If this is a comma-delimited list of directories then the name table is replicated in all of the directories, for redundancy." We mentioned before that we're using a JBOD configuration. We do this even for our NameNode. So in our case the NameNode has two disks mounted at /mnt/disk1 and /mnt/disk2 but you might want to write to just one location if you use RAID. As it says in the documentation the NameNode will write to each of the locations. You can write to a third location: A NFS mount which serves as a backup. Our configuration looks like this:
    <property>
      <name>dfs.name.dir</name>
      <value>/mnt/disk1/hadoop/dfs/name,/mnt/disk2/hadoop/dfs/name</value>
    </property>
    Make sure to create the dfs directories before starting the NameNode. They need to belong to hdfs:hadoop. Formatting the NameNode is all that's left:
    su hdfs -c "/usr/bin/hadoop namenode -format"
    Once you've done all that you can enable the service so it will be started upon system boot and start the NameNode:
    chkconfig hadoop-0.20-namenode on
    service hadoop-0.20-namenode start
    You should be able to see the web interface on your namenode at port 50070 now. Ports that need to be opened to clients on the NameNode are 50070 (web interface, 50470 if you enabled SSL) and 8020 (for HDFS command line interaction). Only port 8020 needs to be enabled for all other servers in the cluster.

    We also use a cron job to run the HDFS Balancer every evening:
    /usr/lib/hadoop-0.20/bin/start-balancer.sh -threshold 5

    2.2 DataNodes
    The DataNodes handle all the data by storing it and serving it to clients. You can run a DataNode on your NameNode and especially for small- or test clusters this is often done but as soon as you have more than three to five machines or rely on your cluster for production use you should use a dedicated NameNode. Setting the DataNodes up is easy though after all our preparations. We need to set the property dfs.data.dir in the file /etc/hadoop/conf/hdfs-site.xml. It "determines where on the local filesystem an DFS data node should store its blocks. If this is a comma-delimited list of directories, then data will be stored in all named directories, typically on different devices. Directories that do not exist are ignored." These are the directories where the real data bytes of HDFS will be written to. If you specify multiple directories the DataNode will write to them in turn which gives good performance when reading the data.

    This is an example of what we are using:
    <property>
      <name>dfs.data.dir</name>
      <value>/mnt/disk1/hadoop/dfs/data,/mnt/disk2/hadoop/dfs/data</value>
    </property>
    Make sure to create the dfs directories before starting the DataNodes. They need to belong to hdfs:hadoop. When that's done you just need to install the DataNode, activate the startup scripts and start it:
    yum install -y hadoop-0.20-datanode
    chkconfig hadoop-0.20-datanode on
    service hadoop-0.20-datanode start
    Your DataNode should be up and running and if you have configured it correctly should also have connected to the NameNode and be visible in the web interface in the Live Nodes list and the configured capacity should go up. Ports that need to be opened to clients are 50075 (web interface, 50475 if you enabled SSL) and 50010 (for data transfer). For the cluster you need to open ports 50010 and 50020.

    3. MapReduce

    MapReduce is split in two parts as well: A JobTracker and multiple TaskTrackers. For small-ish clusters the NameNode and the JobTracker can run on the same server but depending on your usage and available memory you might need to run them on separate servers. We have 18 servers, 17 slaves and 1 master (with NameNode, JobTracker and other services) which isn't a problem so far. We need three properties set on all servers (in mapred-site.xml) to get started.
    • mapred.job.tracker: "The host and port that the MapReduce job tracker runs at. If 'local', then jobs are run in-process as a single map and reduce task."
      • This just points to your JobTracker. There is no default port for this in Hadoop 0.20 but 8021 is often used.
      • Our value (replace $jobtracker with the name or IP of your designated JobTracker): $jobtracker:8021
    • mapred.local.dir: "The local directory where MapReduce stores intermediate data files. May be a comma-separated list of directories on different devices in order to spread disk i/o. Directories that do not exist are ignored."
      • As it says this is a local directory where MapReduce stores stuff an we spread it out over all our discs.
      • Our value: /mnt/disk1/hadoop/mapreduce,/mnt/disk2/hadoop/mapreduce
      • Create the directories on each server with the owner mapred:hadoop
    • mapred.system.dir: "The shared directory where MapReduce stores control files."
      • This is a path in HDFS where MapReduce stores stuff
      • Our value: /hadoop/mapreduce/system
      • If dfs.permissions are on you need to create this directory in HDFS. Execute this command on any server in your cluster: su hdfs -c "/usr/bin/hadoop fs -mkdir /hadoop/mapreduce && /usr/bin/hadoop fs -chown mapred:hadoop /hadoop/mapreduce"

    3.1 JobTracker
    The JobTracker is very easy to setup and start:
    yum install -y hadoop-0.20-jobtracker
    chkconfig hadoop-0.20-jobtracker on
    service hadoop-0.20-jobtracker start
    The web interface should now be available at port 50030 on your JobTracker. Ports 50030 (web interface) and 8021 (not well defined in Hadoop 0.20 but if you followed my configuration this is correct) need to be opened to clients. Only 8021 is necessary for the TaskTrackers.

    If the JobTracker is restarted some old files will not be cleaned up. That's why we added another small cronjob to run daily:
    find /var/log/hadoop/ -type f -mtime +3 -name "job_*_conf.xml" -delete

    3.2 TaskTracker
    The TaskTracker are as easy to install as the JobTracker:
    yum install -y hadoop-0.20-tasktracker
    chkconfig hadoop-0.20-tasktracker on
    service hadoop-0.20-tasktracker start
    The TaskTracker should now be up and running and visible in the JobTracker's Nodes list. Only port 50060 needs to be opened to clients for a minimalistic web interface. Other than that no other ports are needed as TaskTrackers check in at the JobTracker regularly (heartbeat) and get assigned Tasks at the same time.

    4. Configuration

    I'll discuss a few configuration properties here that in a range of "necessary to change" to "nice to know about". I'll mention the following things for each property:
    • The default value,
    • the value we use for our cluster at GBIF if it differs from the default,
    • some of the defaults are quite old and have never been changed so I might mention a value I deem safe to use for everybody,
    • if we set the property to final so it can't be overridden by clients (we set a lot of the parameters to final for purely documentary reasons, even those that can't be overwritten in the first place),
    • if the property has been renamed or deprecated in Hadoop 0.21,
    • and if this property is required in a client configuration file or only on the cluster, if I don't mention it it's not needed.
    Here are the default configuration files for Hadoop 0.20.2 and 0.21:
    And I know that there are some duplications to the section above but I want to keep this Configuration section as a reference.

    core-site.xml

    fs.default.name
    • Default: file:///
    • We: hdfs://$namenode:8020
    • We set this to final
    • Renamed to fs.defaultFS in Hadoop 0.21
    • Needed on the clients
    This is used to specify the default file system and defaults to your local file system that's why it needs be set to a HDFS address. This is important for client configuration as well so your local configuration file should include this element.

    hadoop.tmp.dir
    • Default: /tmp/hadoop-${user.name}
    • CDH3 Default: /var/lib/hadoop-0.20/cache/${user.name}
    • We: Left it at the CDH3 default
    • We set this to final
    As mentioned in the default file this is mainly a base for other temporary directories. If all other configuration options are set correctly there shouldn't be too much data in here.

    fs.trash.interval
    • Default: 0
    • We: 10080
    • We set this to final
    Hadoop has a Trash feature were removed files (using the command line tools) are moved to a .Trash folder in the users home folder. If set to 0 this feature is disabled but if set to a non-zero value this is the amount of minutes between Trash cleaner runs. As we have a lot of users in our system using Hadoop for the first time we chose a safe value here.

    fs.checkpoint.dir
    • Default: ${hadoop.tmp.dir}/dfs/namesecondary
    • We: /mnt/disk1/hadoop/dfs/namesecondary,/mnt/disk2/hadoop/dfs/namesecondary
    • We set this to final
    The secondary NameNode stores its images to merge here. If it is a comma separated list the data is replicated to all these locations on the local disks.

    io.file.buffer.size
    • Default: 4096
    • Safe: 65536
    • We: 131072 (32 * 4096)
    • Can be overwritten by clients
    • We set this to final
    This is used for buffers all over the place to copy, store and write data to. It should be a multiple of 4096 and it should be safe to use 65536 today but we use double that. The performance gain is not enormous but there have been blog posts in the past measuring the impact and it was positive. We've also done our own tests and saw a small performance gain. If you use HBase be careful not to set this too high.

    io.compression.codecs
    • Default: org.apache.hadoop.io.compress.DefaultCodec,org.apache.hadoop.io.compress.GzipCodec,org.apache.hadoop.io.compress.BZip2Codec
    • We: org.apache.hadoop.io.compress.GzipCodec,org.apache.hadoop.io.compress.DefaultCodec,org.apache.hadoop.io.compress.BZip2Codec,com.hadoop.compression.lzo.LzoCodec,com.hadoop.compression.lzo.LzopCodec
    • We set this to final
    This lists all installed compression codecs. If you followed my manual you've got to add two more to the default list of codecs: LzoCodec and LzopCodec.

    io.compression.codec.lzo.class
    • We: com.hadoop.compression.lzo.LzopCodec
    I have actually no idea why this setting is needed as I couldn't find any reference where it is actually used in the code but I didn't look very hard so I might be wrong. All I know is that the documentation mentions that this property needs to be set.

    webinterface.private.actions
    • Default: false
    • We: true
    • We set this to final
    By setting this to true the web interfaces for the JobTracker and NameNode gain some advanced options like killing a job. It makes life a lot easier while still in development or evaluation. But you probably should set this to false once you rely on your Hadoop cluster for production use.

    hdfs-site.xml

    dfs.name.dir
    • Default: ${hadoop.tmp.dir}/dfs/name
    • We: /mnt/disk1/hadoop/dfs/name,/mnt/disk2/hadoop/dfs/name
    • We set this to final
    This is an important setting to set that's why I've already mentioned it above. The NameNode stores stuff in these directories by replicating all information to all these disks. One of them could be a mount on a remote disk (e.g. NFS) to have a backup.
    dfs.data.dir
    • Default: ${hadoop.tmp.dir}/dfs/data
    • We: /mnt/disk1/hadoop/dfs/data,/mnt/disk2/hadoop/dfs/data
    • We set this to final
    This is another important setting as explained above. Different to dfs.name.dir in that the data is not replicated to all disks but distributed among all those locations. The DataNodes save the actual data in these locations. So more space is better. The easiest thing is to use dedicated disks for this. If you save other stuff than Hadoop data on the disks make sure to set dfs.datanode.du.reserved (see below).

    dfs.namenode.handler.count
    • Default: 10
    • We: 20
    • Safe: 10-20
    • We set this to final
    The number of threads the NameNode uses to serve requests. This depends highly on your usage and size of your cluster. We've tried a bunch of different values and settled on 20 without seeing any notable differences. nnbench is probably a good tool to benchmark this. If you've got a large cluster or many file operations (create or delete) you can try upping this value.

    dfs.datanode.handler.count
    • Default: 3
    • We: 5
    • Safe: 5-10
    • We set this to final
    The number of threads DataNodes use. I can't tell what a good value is for large clusters but the TestDFSIO benchmark seems like a good test to run to find a good value here. Just play around. We've tried a bunch of different values up to 20 and didn't see a difference so we chose a value slightly larger than the default.

    dfs.datanode.du.reserved
    • Default: 0
    • We: Left the default
    This many bytes will be left free on the volumes used by the DataNodes (see dfs.data.dir). As our drives are dedicated to Hadoop we left this at 0 but if the drives host other stuff as well set this to an appropriate value.

    dfs.permissions
    • Default: true
    • We: true
    • Renamed to dfs.permissions.enabled in Hadoop 0.21
    • We set this to final
    This enables permission checking in HDFS. Unless you use Secure Hadoop (which we don't that's why I don't cover it here) it is still easy for anyone to read, write and delete anything on the cluster as there is no authentication of users done. So this is purely for safety reasons to avoid messing with the wrong data by accident.

    dfs.replication
    • Default: 3
    • We: 3
    • Can be used in the client configuration
    This is the default replication level used for new files in HDFS. if you change this value later on no existing files will be changed (that can be done on the command line though). Every file in HDFS can have a different replication level. This just sets the default.

    dfs.block.size
    • Default: 67108864
    • We: 134217728
    • Renamed to dfs.blocksize in Hadoop 0.21
    • Can be used in the client configuration
    This factor is on a per file basis and only used for new files. Files saved to HDFS are split in blocks at most this large (64 MB by default). This has multiple implications. The more blocks you have the more load there is on your NameNode. So if you have many files that are larger than the blocksize you might set this larger. If your files are mostly smaller than this you waste no space. All files only take as much space as they actually have data (this is unlike other file systems where a file takes up at least one block no matter how large it really is). So NameNode load (memory requirements as well) are one factor. The deciding factor for us to set this higher per default is that a lot of our calculations in MapReduce are very fast and Mappers finish quickly. As one Mapper usually processes one block and Mappers take a while to set up we chose a higher block size so that each Mapper has more data to process.

    This can be set on a per file basis so you really have to find your own perfect value, perhaps even on a per dataset basis.

    dfs.balance.bandwidthPerSec
    • Default: 1048576
    • We: 2097152
    • Renamed to dfs.datanode.balance.bandwidthPerSec in Hadoop 0.21
    • We set this to final
    This property configures the amounts of bytes per second (default is 1 MB/s) that a DFS balancing operation can use per DataNode. The default is pretty low so we doubled it. We don't use a lot of bandwidth in our cluster at the moment so this is not a problem. Depends on your use case. The higher this number the faster balancing operations will complete. We run balancing every night on a cron job so we want it to be finished by morning.

    dfs.hosts
    • Default: no default set
    • We: /etc/hadoop/conf/allowed_hosts
    • We set this to final
    This file has to contain one name per line. Every name is the name of a DataNode that is allowed to connect to the NameNode. This prevents accidents like what happened to me: I test everything in Virtual Machines so I started a bunch of them, deployed the live config and forgot to change the NameNode so all of a sudden a bunch of Virtual Machines joined our HDFS cluster and blocks began replicating there.... So it is a good thing to explicitly list all allowed hosts in this file.

    dfs.support.append
    • Default: false
    • We: true
    • As far as I know this option has been removed in Hadoop 0.21 and is enabled by default
    • We set this to final
    This option has quite a history. To make it short: If you're using CDH3 set this to true, otherwise leave it false. You want/need this on true if you plan to use HBase.

    dfs.datanode.max.xcievers
    • Default: 256
    • Safe: 1024
    • We: 2048
    • Yes, this is misspelt in Hadoop and it hasn't been fixed in Hadoop 0.21.
    • We set this to final
    This is the maximum number of threads a DataNode may use (for example for file access to the local file system). There used to be bugs in Hadoop so that the default was a bit to low and needed to be set higher. Even today it's worth it to set it higher without a lot of risk. Especially if you're using HBase.

    mapred-site.xml

    HDFS is pretty straightforward to configure and benchmark. MapReduce is more of a black art unfortunately. I'll describe the MapReduce process here because it is important to understand where all the properties come in so you can safely change their values and tweak the performance. In my first draft of this post I wrote that I won't go into much detail on the internals of the MapReduce process. (Un-)fortunately this wasn't as easy as I thought and it has grown into a full blown explanation of everything I know. It is very possible that something's wrong here so please correct me if you see something that is off. And if you're not interested in how this works just skip to the descriptions of the properties itself.

    All of this is valid for Hadoop 0.20.2+737 (the CDH version). I know that some things have changed in Hadoop 0.21 but that's left for another time.

    The Map side

    While a Map is running it is collecting output records in an in-memory buffer called MapOutputBuffer, if there are no reducers a DirectMapOutputCollector is used which makes most of the rest obsolete as it writes immediately to disk. The total size of this in memory buffer is set by the io.sort.mb property and defaults to 100 MB (which is converted to a byte value using a bit shift operation [100 << 20 = 104857600]). Out of these 100 MB io.sort.record.percent are reserved for tracking record boundaries. This property defaults to 0.05 (i.e. 5% which means 5 MB in the default case). Each record to track takes 16 bytes (4 integers of 4 bytes each) of memory which means the buffer can track 327680 map output records with the default settings. The rest of the memory (104857600 bytes - (16 bytes * 327680) = 99614720 bytes) is used to store the actual bytes to be collected (in the default case this will be 95 MB). While Map outputs are collected they are stored in the remaining memory and their location in the in-memory buffer is tracked as well. Once one of these two buffers reaches a threshold specified by io.sort.spill.percent, which defaults to 0.8 (i.e. 80%), the buffer is flushed to disk:
    0.8 * 99614720 = 79691776
    0.8 * 327680 = 262144
    Look in the log output of your Maps and you'll see these three lines at the beginning of every log:
    2010-12-05 01:33:04,912 INFO org.apache.hadoop.mapred.MapTask: io.sort.mb = 100
    2010-12-05 01:33:04,996 INFO org.apache.hadoop.mapred.MapTask: data buffer = 79691776/99614720
    2010-12-05 01:33:04,996 INFO org.apache.hadoop.mapred.MapTask: record buffer = 262144/327680
    You should recognize these numbers!

    Now while the Map is running you might see log lines like these:
    2010-12-05 01:33:08,823 INFO org.apache.hadoop.mapred.MapTask: Spilling map output: record full = true
    2010-12-05 01:33:08,823 INFO org.apache.hadoop.mapred.MapTask: bufstart = 0; bufend = 19361312; bufvoid = 99614720
    2010-12-05 01:33:08,823 INFO org.apache.hadoop.mapred.MapTask: kvstart = 0; kvend = 262144; length = 327680
    2010-12-05 01:33:09,558 INFO org.apache.hadoop.mapred.MapTask: Finished spill 0
    This means we've reached the maximum number of records we can track even though our buffer is still pretty empty (99614720 - 19361312 bytes still free). If however your buffer is the cause of your spill you'll see a line like this:
    2010-12-05 01:33:08,823 INFO org.apache.hadoop.mapred.MapTask: Spilling map output: buffer full = true
    All of this spilling to disk is done in a separate thread so that the Map can continue running. That's also the reason why the spill begins early (when the buffer is only 80% full) so it doesn't fill up before a spill is finished. If one single Map output is too large to fit into the in memory buffer a single spill is done for this one value. A spill actually consists of one file per partition, meaning one file per Reducer.

    After a Map task has finished there may be multiple spills on the TaskTracker. Those files have to be merged into one single sorted file per partition which is then fetched by the Reducers. The property io.sort.factor says how many of those spill files will be merged into one file at a time. The lower the number is the more passes will be required to arrive at the goal. The default is very low and it was considered to set the default to 100 (and in fact looking at the code it sometimes is set to 100 by default). This property can make a pretty huge difference if your Mappers output a lot of data. Not much memory is needed for this property but the larger it is the more open files there will be so make sure to set this to a reasonable value. To find such a value you should run a few MapReduce jobs that you'd expect to see in production use and carefully monitor the log files.

    Watch out for log messages like these:
    • Merging <numSegments> sorted segments
    • Down to the last merge-pass, with <numSegments> segments left of total size: <totalBytes> bytes
    • Merging <segmentsToMerge.size()> intermediate segments out of a total of <totalSegments>
    This is the process on the Map side where this factor is used. If your Mappers only have on spill file all of this doesn't matter. So if you try to benchmark this make sure to use a job with a lot of Map output data. If you only see a line like "Finished spill 0" but none of the above you're only producing one spill file which doesn't require any merging or further sorting. This is the ideal situation and you should try to get the number of spilled records/files as low as possible.

    The Reduce side

    The reduce phase has three different steps: Copy, Sort (which should really be called Merge) and Reduce.

    During the Copy phase the Reducer tries to fetch the output of the Maps from the TaskTrackers and store it on the Reducer either in memory or on disk. The property mapred.reduce.parallel.copies (which defaults to 5) defines how many Threads are started per Reduce task to fetch Map output from the TaskTrackers.

    Here's an example log from the beginning of a Reducer log:
    2010-12-05 01:53:03,846 INFO org.apache.hadoop.mapred.ReduceTask: ShuffleRamManager: MemoryLimit=334063200, MaxSingleShuffleLimit=83515800
    2010-12-05 01:53:03,879 INFO org.apache.hadoop.mapred.ReduceTask: attempt_201012031527_0021_r_000103_0 Need another 1870 map output(s) where 0 is already in progress
    2010-12-05 01:53:03,880 INFO org.apache.hadoop.mapred.ReduceTask: attempt_201012031527_0021_r_000103_0 Thread started: Thread for merging on-disk files
    2010-12-05 01:53:03,880 INFO org.apache.hadoop.mapred.ReduceTask: attempt_201012031527_0021_r_000103_0 Thread waiting: Thread for merging on-disk files
    2010-12-05 01:53:03,880 INFO org.apache.hadoop.mapred.ReduceTask: attempt_201012031527_0021_r_000103_0 Thread started: Thread for merging in memory files
    2010-12-05 01:53:03,881 INFO org.apache.hadoop.mapred.ReduceTask: attempt_201012031527_0021_r_000103_0 Thread started: Thread for polling Map Completion Events
    You can see two things in these log lines. First of all the ShuffleRamManager is started and afterwards you see that this Reducer needs to fetch 1870 map outputs (meaning we had 1870 Mappers). The map output is fetched and shuffled into memory (that's what the ShuffleRamManager is for). You can control its behavior using the mapred.job.shuffle.input.buffer.percent (default is 0.7). Runtime.getRuntime().maxMemory() is used to get the available memory which unfortunately returns slightly incorrect values so be careful when setting this. We'll get back to the last four lines later.

    Our child tasks are running with -Xmx512m (536870912 bytes) so 70% of that should be 375809638 bytes but the ShuffleRamManager reports 334063200. No big deal, just be aware of it. There's a hardcoded limit of 25% of the buffer that a single map output may not surpass. If it is larger than that it will be written to disk (see the MaxSingleShuffleLimit value above: 334063200 * 0.25 = 83515800).

    Now that everything's set up the copiers will start their work and fetch the output. You'll see a bunch of log lines like these:
    2010-12-05 01:53:11,114 INFO org.apache.hadoop.mapred.ReduceTask: header: attempt_201012031527_0021_m_000011_0, compressed len: 454055, decompressed len: 454051
    2010-12-05 01:53:11,114 INFO org.apache.hadoop.mapred.ReduceTask: Shuffling 454051 bytes (454055 raw bytes) into RAM from attempt_201012031527_0021_m_000011_0
    2010-12-05 01:53:11,133 INFO org.apache.hadoop.mapred.ReduceTask: Read 454051 bytes from map-output for attempt_201012031527_0021_m_000011_0
    2010-12-05 01:53:11,133 INFO org.apache.hadoop.mapred.ReduceTask: Rec #1 from attempt_201012031527_0021_m_000011_0 -> (70, 6) from c1n7.gbif.org
    In the first line you see that a map output was successfully copied and it could read the size of the data from the headers. The next line is actually what we've talked about earlier: The map output will now be decompressed (if it was compressed) and saved into memory using the ShuffleRamManager. The third line acknowledges that this succeeded. And the last line is information for a bug and should have been removed already according to a comment in the source code.

    If for whatever reason the map output doesn't fit into memory you will see a similar log line to the second one above but "RAM" will be replaced by "Local-FS" and the fourth line will be missing. You obviously want as much data into memory as possible so shuffling on to the Local-FS is a warning sign or at least a sign for possible optimizations.

    While all this goes on until all map outputs have been fetched there are two threads (Thread for merging on-disk files and Thread for merging in memory files) waiting for some conditions until they become active. The conditions are as follows:
    • The used memory in the in-memory buffer is above mapred.job.shuffle.merge.percent (default ist 66%, in our example that would mean 334063200 * 0.66 = 220481712 bytes) and there are at least two map outputs in the buffer
    • or there are more than mapred.inmem.merge.threshold (defaults to 1000) map outputs in the in-memory buffer, independent of the size
    • or if there are more than io.sort.factor * 2 -1 files on disk.
    When one of the first two condition triggers you'll see something like this:
    2010-12-05 01:53:42,106 INFO org.apache.hadoop.mapred.ReduceTask: Initiating in-memory merge with 501 segments...
    2010-12-05 01:53:42,114 INFO org.apache.hadoop.mapred.Merger: Merging 501 sorted segments
    ...
    2010-12-05 01:53:46,492 INFO org.apache.hadoop.mapred.ReduceTask: attempt_201012031527_0021_r_000103_0 Merge of the 501 files in-memory complete. Local file is /mnt/disk1/hadoop/mapreduce/local/taskTracker/lfrancke/jobcache/job_201012031527_0021/attempt_201012031527_0021_r_000103_0/output/map_1.out of size 220545981
    2
    This could actually trigger the third condition as it writes a new file to disk. When that happens you'll see something like this:
    2010-12-10 14:28:23,289 INFO org.apache.hadoop.mapred.ReduceTask: attempt_201012101346_0001_r_000012_0We have  19 map outputs on disk. Triggering merge of 10 files
    The io.sort.factor was set to the default of 10. 10 (out of the 19) files will be merged into one, leaving 10 on disk (i.e. io.sort.factor).

    Both of these (the in-memory and the on-disk merge, the latter is also called Interleaved on-disk merge) will produce a new single output file and write it to disk. All of this is only going on as long as map outputs are still fetched. When that's finished we wait for running merges to finish but won't start any new ones in these threads:
    2010-12-05 01:59:10,598 INFO org.apache.hadoop.mapred.ReduceTask: GetMapEventsThread exiting
    2010-12-05 01:59:10,599 INFO org.apache.hadoop.mapred.ReduceTask: getMapsEventsThread joined.
    2010-12-05 01:59:10,599 INFO org.apache.hadoop.mapred.ReduceTask: Closed ram manager
    2010-12-05 01:59:10,599 INFO org.apache.hadoop.mapred.ReduceTask: Interleaved on-disk merge complete: 3 files left.
    2010-12-05 01:59:10,599 INFO org.apache.hadoop.mapred.ReduceTask: In-memory merge complete: 314 files left.
    As you can see by the timestamps no merges were running in our case so everything just shut down. During the copy phase we finished a total of three in-memory merges that's why we currently have three files on the disk. 314 more map outputs are still in the in-memory buffer. This concludes the Copy phase and the Sort phase begins:
    2010-12-05 01:59:10,605 INFO org.apache.hadoop.mapred.Merger: Merging 314 sorted segments
    2010-12-05 01:59:10,605 INFO org.apache.hadoop.mapred.Merger: Down to the last merge-pass, with 314 segments left of total size: 127512782 bytes
    2010-12-05 01:59:13,903 INFO org.apache.hadoop.mapred.ReduceTask: Merged 314 segments, 127512782 bytes to disk to satisfy reduce memory limit
    2010-12-05 01:59:13,904 INFO org.apache.hadoop.mapred.ReduceTask: Merging 4 files, 788519164 bytes from disk
    2010-12-05 01:59:13,905 INFO org.apache.hadoop.mapred.ReduceTask: Merging 0 segments, 0 bytes from memory into reduce
    2010-12-05 01:59:13,905 INFO org.apache.hadoop.mapred.Merger: Merging 4 sorted segments
    2010-12-05 01:59:14,493 INFO org.apache.hadoop.mapred.Merger: Down to the last merge-pass, with 4 segments left of total size: 788519148 bytes
    There are two things happening here. First of all the remaining 314 files that are still in memory are merged into one file on the disk (the first three lines). So now there are four files on the disk. These four files are merged into one.

    There is an option mapred.job.reduce.input.buffer.percent which is set to 0 by default which allows the Reducer to keep some map output files in memory. The following is a snippet with this property set to 0.7:
    2010-12-05 23:11:55,657 INFO org.apache.hadoop.mapred.ReduceTask: Merging 3 files, 661137901 bytes from disk
    2010-12-05 23:11:55,660 INFO org.apache.hadoop.mapred.ReduceTask: Merging 312 segments, 127381881 bytes from memory into reduce
    2010-12-05 23:11:55,661 INFO org.apache.hadoop.mapred.Merger: Merging 3 sorted segments
    2010-12-05 23:11:55,688 INFO org.apache.hadoop.mapred.Merger: Down to the last merge-pass, with 3 segments left of total size: 661137889 bytes
    2010-12-05 23:11:55,688 INFO org.apache.hadoop.mapred.Merger: Merging 313 sorted segments
    2010-12-05 23:11:55,689 INFO org.apache.hadoop.mapred.Merger: Down to the last merge-pass, with 313 segments left of total size: 788519778 bytes
    You can see that instead of merging the 312 segments from memory to disk they are kept in memory while the three files on disk are merged into one and all of the resulting 313 segments are streamed into the reducer.

    There seems to be a bug in Hadoop though. I'm not 100% sure about this one so any insight would be appreciated. When the following conditions are true segments from the memory don't seem to be written to disk even if they should be according to the configuration:
    • There are segments in memory that should be written to disk before the reduce task begins according to mapred.job.reduce.input.buffer.percent
    • and there are more files on disk than io.sort.factor
    If this happens you see this:
    2010-12-10 16:39:40,671 INFO org.apache.hadoop.mapred.ReduceTask: Keeping 14 segments, 18888592 bytes in memory for intermediate, on-disk merge
    2010-12-10 16:39:40,673 INFO org.apache.hadoop.mapred.ReduceTask: Merging 10 files, 4143441520 bytes from disk
    2010-12-10 16:39:40,674 INFO org.apache.hadoop.mapred.ReduceTask: Merging 0 segments, 0 bytes from memory into reduce
    2010-12-10 16:39:40,674 INFO org.apache.hadoop.mapred.Merger: Merging 24 sorted segments
    2010-12-10 16:39:40,859 INFO org.apache.hadoop.mapred.Merger: Down to the last merge-pass, with 24 segments left of total size: 4143441480 bytes
    So the steps being done in the Sort phase are the following:
    1. Merge all segments (= map outputs) that are still in memory and don't fit into the memory specified by mapred.job.reduce.input.buffer.percent into one file on disk if there are less than io.sort.factor files on disk so we end up with at most io.sort.factor files on the disk after this step. If there are already io.sort.factor or more files on disk but there are map outputs that need to be written out of memory keep them in memory for now
      1. In the first case you'll see a log message like this: Merged 314 segments, 127512782 bytes to disk to satisfy reduce memory limit
      2. In the second case you'll see this: Keeping 14 segments, 18888592 bytes in memory for intermediate, on-disk merge
    2. All files on disk and all remaining files in memory that need to be merged (case 1.b) are determined. You'll see a log message like this: "Merging 4 files, 788519164 bytes from disk".
    3. All files that remain in memory during the Reduce phase are determined: "Merging 312 segments, 127381881 bytes from memory into reduce".
    4. All files (on disk + in-memory) from step 2. are merged together using io.sort.factor as the merge factor. Which means that there might be intermediate merges to disk.
    5. Merge all remaining in-memory (from step 3.) and on-disk files (from step 4.) into one stream to be read by the Reducer. This is done in a streaming fashion without writing new data to disk and just returning an Iterator to the Reduce phase.
    This Iterator is given to the Reducer and so the Reduce phase starts.

    Well, it turned out to be a rather detailed description of the process which is helpful to understand the configuration properties available to you. See below for a detailed list of all the relevant properties:
    io.sort.factor
    • Default: 10
    • We: 100
    • Safe: 20-100
    • Renamed to mapreduce.task.io.sort.factor in Hadoop 0.21
    • Can be used in the client configuration
    I've explained pretty thoroughly what this parameter does so I won't go into detail here. The whole situation with io.sort.factor and io.sort.mb is not ideal but as long as they are the options we have and the defaults are very low it is pretty safe to change them to a more reasonable value. It is worthwhile to take a look at your logs and search for the lines mentioned in the explanation above. This can be set on a per-job basis and for jobs that run frequently it's worth to find a good job specific value.

    io.sort.mb
    • Default: 100
    • Renamed to mapreduce.task.io.sort.mb in Hadoop 0.21
    • Can be used in the client configuration
    You can adjust the amount of memory used in the Mappers to collect Map outputs with this parameter. This parameter obviously depends heavily on the amount of memory you have available in total for your child VMs and on the memory requirements of your tasks. Your goal should be to minimize the amount of spilling that has to be performed as explained above and to utilize the available as best as possible. If your Map tasks don't need a lot of memory themselves you can use almost all available memory here. The default settings allocate 200 MB for child VMs and half of that is used for the output buffer so your Map tasks has about 100 MB available by default.

    io.sort.record.percent
    • Default: 0.05
    • This has been removed in favor of automatic configuration in Hadoop 0.21
    • Can be used in the client configuration
    The output buffer on the map side is split in two parts. One stores the actual bytes of the output data and the other one stores 16 bytes of metadata per output. This property specifies how much memory of the buffer (io.sort.mb) is used for tracking the metadata. The default is 5% and is often very low for jobs which output only small amounts of data in their map tasks. Look for lines indicating whether a spill to disk occurs because of record full = true. If this happens try to increase this value. This is another property which is very specific to the jobs you're running so it might need tuning for each and every job. Thankfully this mechanism has been replaced in Hadoop 0.21.

    io.sort.spill.percent
    • Default: 0.8
    • Renamed to mapreduce.map.sort.spill.percent in Hadoop 0.21
    • Can be used in the client configuration
    This property just configures when the data from the map output buffer will be written (spilled) to disk. The spilling process is running in a separate thread and output will be collected while it is running so it is important to start this process before the buffer is completely full as the map tasks will pause until there is space available.

    mapred.job.tracker
    • Default: local
    • We: <jobtracker>:8021
    • We set this to final
    • Renamed to mapreduce.jobtracker.address in Hadoop 0.21
    • Needed on the clients
    This lets the client know where to find the JobTracker and it lets the JobTracker know which port to bind to.

    mapred.local.dir
    • Default: ${hadoop.tmp.dir}/mapred/local
    • We: /mnt/disk1/hadoop/mapreduce/local,/mnt/disk2/hadoop/mapreduce/local
    • We set this to final
    • Renamed to mapreduce.cluster.local.dir in Hadoop 0.21
    This lets the MapReduce servers know where to store intermediate files. This may be a comma-separated list of directories to spread the load. Make sure there's enough space here for all your intermediate files. We share the same disks for MapReduce and HDFS.

    mapred.system.dir
    • Default: ${hadoop.tmp.dir}/mapred/system
    • We: /hadoop/mapred/system
    • We set this to final
    • Renamed to mapreduce.jobtracker.system.dir in Hadoop 0.21
    This is a folder in the defaultFS where MapReduce stores some control files. In our case that would be a directory in HDFS. If you have dfs.permissions (which it is by default) enabled make sure that this directory exists and is owned by mapred:hadoop.

    mapred.temp.dir
    • Default: ${hadoop.tmp.dir}/mapred/temp
    • We: /tmp/mapreduce
    • We set this to final
    • Renamed to mapreduce.cluster.temp.dir in Hadoop 0.21
    This is a folder to store temporary files in. It is hardly - if at all used. If I understand the description correctly this is supposed to be in HDFS but I'm not entirely sure by reading the source code. So we set this to a directory that exists on the local filesystem as well as in HDFS.

    mapred.map.tasks
    • Default: 2
    • Renamed to mapreduce.job.maps in Hadoop 0.21
    It is important to realize that this is just a hint for MapReduce as to the number of Maps it should use. In most cases this value is ignored and the actual number of Maps is dependent on the input data and generated automatically. For those rare cases where this value is used we set it to about 90% of our map slot capacity. This can be set client-side per job so if you have a job that relies on this property you better set it there to an appropriate value.

    mapred.reduce.tasks
    • Default: 1
    • Renamed to mapreduce.job.reduces in Hadoop 0.21
    This is different than the property for map tasks in that it is often not possible to calculate a "native" or optimal number of reduce tasks for a job. With this property you can specify the number of reduce tasks to start for a given job. The default is very low. The description suggests to set this to 99% of the cluster capacity so that all reduces finish in one wave. This is sensible when you use the default scheduler but as soon as multiple jobs run in parallel it's hard to guarantee that all reduces of one job finish in one wave. We're constantly playing around with this and currently have this at about 50% of our capacity.

    This too can be specified on a per-job basis.

    mapred.jobtracker.taskScheduler
    • Default: org.apache.hadoop.mapred.JobQueueTaskScheduler
    • We: org.apache.hadoop.mapred.FairScheduler
    • We set this to final
    • Renamed to mapreduce.jobtracker.taskscheduler in Hadoop 0.21
    With the default configuration all jobs are placed in a priority FIFO queue and submitted one after the other. This is fine for testing but it doesn't utilize the available resources very well. This property allows you to change the scheduler used. These are the available schedulers in CDH3b3:
    Depending on the scheduler you decide to use there may be additional properties which I'm not going to mention here. Have a look at the dedicated documentation.

    mapred.reduce.parallel.copies
    • Default: 5
    • We: ~20-50
    • We set this to final
    • Renamed to mapreduce.reduce.shuffle.parallelcopies in Hadoop 0.21
    The reduce tasks have to fetch the map outputs from the remote servers. They have to fetch the output from each map of which there may be thousands. This option allows to parallelize the copy process. Tuning this to a value is very worthwhile. In our first tests this property gave us one of the best performance increases of all properties. We started to increase this property in steps of 5 and looked very carefully at the logs and our monitoring system to find a value that works for us. We've not yet finished this process but values between 20 and 50 seem to mostly work without problems.

    mapred.tasktracker.map.tasks.maximum & mapred.tasktracker.reduce.tasks.maximum
    • Default: 2
    • We set this to final
    • Renamed to mapreduce.tasktracker.map.tasks.maximum & mapreduce.tasktracker.map.tasks.maximum in Hadoop 0.21
    This setting is very important and we've yet to find values that we are comfortable with. This setting can be different on each TaskTracker and defines how many map or reduce task "slots" there are on a specific TaskTracker. You need to set these to values that don't overload your servers while still fully utilizing them. You also need to make sure that there's enough memory for all tasks and services running on a server (see mapred.child.java.opts). By setting this property to different values depending on your server configuration you can easily use heterogeneous hardware in your cluster. Each distinct hardware configuration will have these properties set to different values. A general rule from the Hadoop Definitive Guide book says that these properties can be set to number of cores - 1. We've tried various settings now but found the load on the servers to be very high with those settings so we'll have to do more benchmarking.

    mapred.child.java.opts
    • Default: -Xmx200m
    • Can be used in the client configuration
    These are the options given to each child JVM started (map- and reduce tasks). The default just sets the maximum memory to 200 MB. This can be set on the client to pass options needed for a specific job. GC logging for example can be enabled as well. This isn't configurable on a per TaskTracker basis so you have to make sure that every machine in your cluster fulfills the requirements. Available memory needs to be at least (mapred.tasktracker.map.tasks.maximum + mapred.tasktracker.reduce.tasks.maximum) * Xmx.

    mapred.inmem.merge.threshold
    • Default: 1000
    • We: 0
    • Renamed to mapreduce.reduce.merge.inmem.threshold in Hadoop 0.21
    • Can be used in the client configuration
    I've explained the effect of this property in the MapReduce description above but to reiterate: The reduce side fetches map outputs to memory. Once the memory is full or this many map outputs are in memory they are merged together to one file on the disk. This can be set on a per job basis but as a default we've disabled this behavior and just flush to disk when the memory is full. This seems to have been better for all our jobs so far but it's definitely a property to look out for.

    mapred.job.shuffle.merge.percent
    • Default: 0.66
    • Renamed to mapreduce.reduce.shuffle.merge.percent in Hadoop 0.21
    • Can be used in the client configuration
    Once the memory buffer in the copy (shuffle) phase of the reduce task is this full a background thread will start to merge all map outputs collected in memory so far and write them to a single file on disk. This is similar to what's happening on the map side. In the default configuration the mapred.inmem.merge.threshold parameter might actually trigger a merge before this value is hit. We haven't yet played around with this property but you'd have to be careful to turn it not too high so that the copy processes have to wait for the buffer to be empty again. That could be a huge performance hit.

    An addition to Hadoop's logging would be nice that lets us know how full the buffer is the moment a merge finishes.

    mapred.job.shuffle.input.buffer.percent
    • Default: 0.7
    • Renamed to mapreduce.reduce.shuffle.input.buffer.percent in Hadoop 0.21
    • Can be used in the client configuration
    This is the amount of memory from the total available memory (specified by mapred.child.java.opts) that's allocated for collecting map outputs in memory on the reduce side. Another parameter we haven't played around with but my guess would be that this can be easily set a little bit higher.

    mapred.job.reduce.input.buffer.percent
    • Default: 0.0
    • Renamed to mapreduce.reduce.input.buffer.percent in Hadoop 0.21
    • Can be used in the client configuration
    Usually map outputs would be written to disk when the sort phase (on the reduce) ends. If you have reduce tasks that don't need a lot of memory themselves you can set this to a higher value so that map outputs up to this amount of memory (in percent of the total available memory) aren't written to disk but kept in memory. This is obviously faster than an intermediate spill to disk. Should be considered on a per-job basis.

    mapred.map.tasks.speculative.execution & mapred.reduce.tasks.speculative.execution
    • Default: true
    • We: false
    • We will set this to final once we're in production
    • Renamed to mapreduce.map.speculative & mapreduce.reduce.speculative in Hadoop 0.21
    • Can be used in the client configuration
    Speculative Execution starts multiple instances of certain map or reduce tasks when it detects certain circumstances (like an unusually slow task or node) to avoid waiting for stragglers too long. This sounds like a good idea and we've got it enabled at the moment but when we go to production this will probably be disabled as it uses valuable resources on the cluster that mostly goes to waste and while one job may finish faster all the others have to wait longer.

    mapred.job.reuse.jvm.num.tasks
    • Default: 1
    • Renamed to mapreduce.job.jvm.numtasks in Hadoop 0.21
    • Can be used in the client configuration
    Child JVMs are spawned for the map and reduce tasks. This parameter lets you reuse these VMs for multiple tasks. The default value creates a new JVM for each task which has some overhead (the book says about one second per JVM). We've played around with it a bit and it can make things faster but you've got to be careful with memory leaks and shared state. Basically you should be sure that your jobs can handle this. If you have a performance critical job you can play around with this but we've had some OutOfMemory errors when using this so we're conservative at the moment. If you set it to -1 a JVM will never be destroyed.

    tasktracker.http.threads
    • Default: 40
    • We: 80
    • We set this to final
    • Renamed to mapreduce.tasktracker.http.threads in Hadoop 0.21
    The map output is fetched by the reducers from the TaskTrackers via HTTP. This property lets you adjust the number of threads that server those requests. When we upped the parallel copies we had some errors about fetch-failures so we slowly increased this value. Those two parameters need to be carefully tuned. 80 seemed to cause no problems for us so we stuck to it for now. You have to restart your TaskTrackers after changing this value.

    mapred.compress.map.output
    • Default: false
    • We: true
    • Renamed to mapreduce.map.output.compress in Hadoop 0.21
    • Can be used in the client configuration
    Turning this on will compress the output of your Mappers using SequenceFile compression. Depending on the codec you chose this computation may be CPU intensive and result in varying degrees of compression. We've benchmarked jobs of different sizes with this intermediate compression enabled and disabled and while some of them took slightly longer than before it is still good to enable it. The cost isn't too high and there is a lot less intermediate data generated. Less I/O in general is good especially if multiple jobs are running.

    mapred.map.output.compression.codec
    • Default: org.apache.hadoop.io.compress.DefaultCodec
    • We: com.hadoop.compression.lzo.LzoCodec
    • Renamed to mapreduce.map.output.compress.codec in Hadoop 0.21
    • Can be used in the client configuration
    With this property you specify the specific compression codec to use for Map output compression. So far we've only tried LZO. This choice was based on the experience of others and the general properties of the algorithm being very fast but sacrificing a bit of compression efficiency for its speed. We plan to test the other algorithms as well.

    mapred.hosts
    • Default: no default set
    • We: /etc/hadoop/conf/allowed_hosts
    • Renamed to mapreduce.jobtracker.hosts.filename in Hadoop 0.21
    • We set this to final
    This is the same as dfs.hosts just specifies which TaskTrackers are allowed to get work from the JobTracker. They both have the same format so it's quite common for them to be the same file.

    Conclusion

    After setting up all these parameters the way you like them you should have a fully functional but basic Hadoop cluster running. You can submit jobs, use HDFS etc. But there are a few more things that we can do like installing Hive, Hue, Pig, Sqoop, etc. We've also yet to cover Puppet. All of this is hopefully forthcoming in more blog posts in the future.

    We're also very interested in other users (or interested people and companies) of Hadoop, HBase & Co. in Scandinavia who would be interested in a Hadoop Meetup. We're located in Copenhagen. Contact us if you're interested.

    If you have any questions or spot any problems or mistakes please let me know in the comments or by mail.