Hi Stanley, Evaluating and analysing benchmark results of a system is complicated, it requires some deep understanding of the target system. Probably you can get some hints from papers about MapReduce/Hadoop, in which the experiment section typically provides good examples on how to evaluate test results. Benchmakring work for other systems (e.g., file systems) is probably helpful for you as well.
Wantao ------------------ Original ------------------ From: "Marcos Ortiz"<[email protected]>; Date: Sun, May 8, 2011 02:14 PM To: "stanley.shi"<[email protected]>; Cc: "mapreduce-user"<[email protected]>; Subject: Re: FW: NNbench and MRBench El 5/8/2011 12:46 AM, [email protected] escribió: Thanks Marcos. This post of Michael Noll does provide some information about how to run these benchmarks, but there's not much information about how to evaluate the results. Do you know some resources about the result analysis? Thanks very much :) Regards, Stanley -----Original Message----- From: Marcos Ortiz [mailto:[email protected]] Sent: 2011年5月8日 11:09 To: [email protected] Cc: Shi, Stanley Subject: Re: FW: NNbench and MRBench El 5/7/2011 10:33 PM, [email protected] escribió: Thanks, Marcos, Through these links, I still can't find anything about the NNbench and MRBench. -----Original Message----- From: Marcos Ortiz [mailto:[email protected]] Sent: 2011年5月8日 10:23 To: [email protected] Cc: Shi, Stanley Subject: Re: FW: NNbench and MRBench El 5/7/2011 8:53 PM, [email protected] escribió: Hi guys, I have a cluster of 16 machines running Hadoop. Now I want to do some benchmark on this cluster with the "nnbench" and "mrbench". I'm new to the hadoop thing and have no one to refer to. I don't know what the supposed result should I have? Now for mrbench, I have an average time of 22sec for a one map job. Is this too bad? What the supposed results might be? For nnbench, what's the supposed results? Below is my result. ================ Date& time: 2011-05-05 20:40:25,459 Test Operation: rename Start time: 2011-05-05 20:40:03,820 Maps to run: 1 Reduces to run: 1 Block Size (bytes): 1 Bytes to write: 0 Bytes per checksum: 1 Number of files: 10000 Replication factor: 1 Successful file operations: 10000 # maps that missed the barrier: 0 # exceptions: 0 TPS: Rename: 1763 Avg Exec time (ms): Rename: 0.5672 Avg Lat (ms): Rename: 0.4844 null RAW DATA: AL Total #1: 4844 RAW DATA: AL Total #2: 0 RAW DATA: TPS Total (ms): 5672 RAW DATA: Longest Map Time (ms): 5672.0 RAW DATA: Late maps: 0 RAW DATA: # of exceptions: 0 ============================= One more question, when I set maps number to bigger, I get all zeros results: ============================= Test Operation: create_write Start time: 2011-05-03 23:22:39,239 Maps to run: 160 Reduces to run: 160 Block Size (bytes): 1 Bytes to write: 0 Bytes per checksum: 1 Number of files: 1 Replication factor: 1 Successful file operations: 0 # maps that missed the barrier: 0 # exceptions: 0 TPS: Create/Write/Close: 0 Avg exec time (ms): Create/Write/Close: 0.0 Avg Lat (ms): Create/Write: NaN Avg Lat (ms): Close: NaN RAW DATA: AL Total #1: 0 RAW DATA: AL Total #2: 0 RAW DATA: TPS Total (ms): 0 RAW DATA: Longest Map Time (ms): 0.0 RAW DATA: Late maps: 0 RAW DATA: # of exceptions: 0 ===================== Can anyone point me to some documents? I really appreciate your help :) Thanks, stanley You can use these resources: http://www.michael-noll.com/blog/2011/04/09/benchmarking-and-stress-testing-an-hadoop-cluster-with-terasort-testdfsio-nnbench-mrbench/ http://answers.oreilly.com/topic/460-how-to-benchmark-a-hadoop-cluster/ http://wiki.apache.org/hadoop/HardwareBenchmarks http://www.quora.com/Apache-Hadoop/Are-there-any-good-Hadoop-benchmark-problems Regards Well, on the Micheal Noll's post says this: NameNode benchmark (nnbench) ======================= NNBench (see src/test/org/apache/hadoop/hdfs/NNBench.java) is useful for load testing the NameNode hardware and configuration. It generates a lot of HDFS-related requests with normally very small "payloads" for the sole purpose of putting a high HDFS management stress on the NameNode. The benchmark can simulate requests for creating, reading, renaming and deleting files on HDFS. I like to run this test simultaneously from several machines -- e.g. from a set of DataNode boxes -- in order to hit the NameNode from multiple locations at the same time. The syntax of NNBench is as follows: NameNode Benchmark 0.4 Usage: nnbench <options> Options: -operation <Available operations are create_write open_read rename delete. This option is mandatory> * NOTE: The open_read, rename and delete operations assume that the files they operate on, are already available. The create_write operation must be run before running the other operations. -maps <number of maps. default is 1. This is not mandatory> -reduces <number of reduces. default is 1. This is not mandatory> -startTime <time to start, given in seconds from the epoch. Make sure this is far enough into the future, so all maps (operations) will start at the same time>. default is launch time + 2 mins. This is not mandatory -blockSize <Block size in bytes. default is 1. This is not mandatory> -bytesToWrite <Bytes to write. default is 0. This is not mandatory> -bytesPerChecksum <Bytes per checksum for the files. default is 1. This is not mandatory> -numberOfFiles <number of files to create. default is 1. This is not mandatory> -replicationFactorPerFile <Replication factor for the files. default is 1. This is not mandatory> -baseDir <base DFS path. default is /becnhmarks/NNBench. This is not mandatory> -readFileAfterOpen <true or false. if true, it reads the file and reports the average time to read. This is valid with the open_read operation. default is false. This is not mandatory> -help: Display the help statement The following command will run a NameNode benchmark that creates 1000 files using 12 maps and 6 reducers. It uses a custom output directory based on the machine's short hostname. This is a simple trick to ensure that one box does not accidentally write into the same output directory of another box running NNBench at the same time. $ hadoop jar hadoop-*-test.jar nnbench -operation create_write \ -maps 12 -reduces 6 -blockSize 1 -bytesToWrite 0 -numberOfFiles 1000 \ -replicationFactorPerFile 3 -readFileAfterOpen true \ -baseDir /benchmarks/NNBench-`hostname -s` Note that by default the benchmark waits 2 minutes before it actually starts! MapReduce benchmark (mrbench) ======================= MRBench (see src/test/org/apache/hadoop/mapred/MRBench.java) loops a small job a number of times. As such it is a very complimentary benchmark to the "large-scale" TeraSort benchmark suite because MRBench checks whether small job runs are responsive and running efficiently on your cluster. It puts its focus on the MapReduce layer as its impact on the HDFS layer is very limited. This test should be run from a single box (see caveat below). The command syntax can be displayed via mrbench --help: MRBenchmark.0.0.2 Usage: mrbench [-baseDir ] [-jar ] [-numRuns ] [-maps ] [-reduces ] [-inputLines ] [-inputType ] [-verbose] Important note: In Hadoop 0.20.2, setting the -baseDir parameter has no effect. This means that multiple parallel MRBench runs (e.g. started from different boxes) might interfere with each other. This is a known bug (MAPREDUCE-2398). I have submitted a patch but it has not been integrated yet. In Hadoop 0.20.2, the parameters default to: -baseDir: /benchmarks/MRBench [*** see my note above ***] -numRuns: 1 -maps: 2 -reduces: 1 -inputLines: 1 -inputType: ascending The command to run a loop of 50 small test jobs is: $ hadoop jar hadoop-*-test.jar mrbench -numRuns 50 Exemplary output of the above command: DataLines Maps Reduces AvgTime (milliseconds) 1 2 1 31414 This means that the average finish time of executed jobs was 31 seconds. Can you check this? http://www.slideshare.net/ydn/ahis2011-platform-hadoop-simulation-and-performance http://issues.apache.org/jira/browse/HADOOP-5867 Did you search on the current documentation of the API? Regards Ok, I understand. Let me try to help you, because I'm a newie on the Hadoop ecosystem. Tom White on its answer to this topic on the OReilly Answers's Site does a introduction to this: The following command writes 10 files of 1,000 MB each: % hadoop jar $HADOOP_INSTALL/hadoop-*-test.jar TestDFSIO -write -nrFiles 10 -fileSize 1000 At the end of the run, the results are written to the console and also recorded in a local file (which is appended to, so you can rerun the benchmark and not lose old results): % cat TestDFSIO_results.log ----- TestDFSIO ----- : write Date & time: Sun Apr 12 07:14:09 EDT 2009 Number of files: 10 Total MBytes processed: 10000 Throughput mb/sec: 7.796340865378244 Average IO rate mb/sec: 7.8862199783325195 IO rate std deviation: 0.9101254683525547 Test exec time sec: 163.387 The files are written under the /benchmarks/TestDFSIO directory by default (this can be changed by setting the test.build.data system property), in a directory called io_data. To run a read benchmark, use the -read argument. Note that these files must already exist (having been written by TestDFSIO -write): % hadoop jar $HADOOP_INSTALL/hadoop-*-test.jar TestDFSIO -read -nrFiles 10 -fileSize 1000 Here are the results for a real run: ----- TestDFSIO ----- : read Date & time: Sun Apr 12 07:24:28 EDT 2009 Number of files: 10 Total MBytes processed: 10000 Throughput mb/sec: 80.25553361904304 Average IO rate mb/sec: 98.6801528930664 IO rate std deviation: 36.63507598174921 ----------------------------------------- Test exec time sec: 47.624 When you’ve finished benchmarking, you can delete all the generated files from HDFS using the -clean argument: % hadoop jar $HADOOP_INSTALL/hadoop-*-test.jar TestDFSIO -clean You can see that all results are written to the TestDFSIO_results.log. So, you can begin to experiment with this. You can continue this reading on the Chapter 9 of the Hadoop: The Definitive Guide 2nd Edition, on the topic: Benchmarking a Hadoop Cluster. In it, Tom gives several advices to benchmark a Hadoop Cluster: - Use a cluster that is not been used by others - One of the primary test that one should do is a intensive I/O benchmark, to prove the cluster before it goes live to production - Write benchmarks with Gridmix (Check this http://developer.yahoo.net/blogs/hadoop/2010/04/gridmix3_emulating_production.html) Well, I hope that this information could help you. Remember, I've worked with Hadoop only for 1 year, so, you can ask for advices to others colleagues too. Regards -- Marcos Luís Ortíz Valmaseda Software Engineer (Large-Scaled Distributed Systems) University of Information Sciences, La Habana, Cuba Linux User # 418229 http://about.me/marcosortiz
