I just finished some testing with JDK 1.6 u17 - so far no performance improvements with just changing that. Disabling LZO compression did gain a little bit (up to about 30/sec from 25/sec per thread). Turning of indexes helped the most - that brought me up to 115/sec @ 2875 total rows a second. A single perl/thrift process can load at over 350 rows/sec so its not scaling as well as I would have expected, even without the indexes.
Are the transactional indexes that costly? What is the bottleneck there? CPU utilization and network packets went up when I disabled the indexes, I don't think those are the bottlenecks for the indexes. I was even able to add another 15 insert process (total of 40) and only lost about 10% on a per process throughput. I probably could go even higher, none of the nodes are above CPU 60% utilization and IO wait was at most 3.5%. Each rowkey is unique, so there should not be any blocking on the row locks. I'll do more indexed tests tomorrow. thanks, -chris On Apr 29, 2010, at 12:18 AM, Todd Lipcon wrote: > Definitely smells like JDK 1.6.0_18. Downgrade that back to 16 or 17 and you > should be good to go. _18 is a botched release if I ever saw one. > > -Todd > > On Wed, Apr 28, 2010 at 10:54 PM, Chris Tarnas <c...@email.com> wrote: > >> Hi Stack, >> >> Thanks for looking. I checked the ganglia charts, no server was at more >> than ~20% CPU utilization at any time during the load test and swap was >> never used. Network traffic was light - just running a count through hbase >> shell generates a much higher use. One the server hosting meta specifically, >> it was at about 15-20% CPU, and IO wait never went above 3%, was usually >> down at near 0. >> >> The load also died with a thrift timeout on every single node (each node >> connecting to localhost for its thrift server), it looks like a datanode >> just died and caused every thrift connection to timeout - I'll have to up >> that limit to handle a node death. >> >> Checking logs this appears in the logs of the region server hosting meta, >> looks like the dead datanode causing this error: >> >> 2010-04-29 01:01:38,948 WARN org.apache.hadoop.hdfs.DFSClient: >> DFSOutputStream ResponseProcessor exception for block >> blk_508630839844593817_11180java.io.IOException: Bad response 1 for block >> blk_508630839844593817_11180 from datanode 10.195.150.255:50010 >> at >> org.apache.hadoop.hdfs.DFSClient$DFSOutputStream$ResponseProcessor.run(DFSClient.java:2423) >> >> The regionserver log on teh dead node, 10.195.150.255 has some more errors >> in it: >> >> http://pastebin.com/EFH9jz0w >> >> I found this in the .out file on the datanode: >> >> # Java VM: Java HotSpot(TM) 64-Bit Server VM (16.0-b13 mixed mode >> linux-amd64 ) >> # Problematic frame: >> # V [libjvm.so+0x62263c] >> # >> # An error report file with more information is saved as: >> # /usr/local/hadoop-0.20.1/hs_err_pid1364.log >> # >> # If you would like to submit a bug report, please visit: >> # http://java.sun.com/webapps/bugreport/crash.jsp >> # >> >> >> There is not a single error in the datanode's log though. Also of note - >> this happened well into the test, so the node dying cause the load to abort >> but not the prior poor performance. Looking through the mailing list it >> looks like java 1.6.0_18 has a bad rep so I'll update the AMI (although I'm >> using the same JVM on other servers in the office w/o issue and decent >> single node performance and never dying...). >> >> Thanks for any help! >> -chris >> >> >> >> >> On Apr 28, 2010, at 10:10 PM, Stack wrote: >> >>> What is load on the server hosting meta like? Higher than others? >>> >>> >>> >>> On Apr 28, 2010, at 8:42 PM, Chris Tarnas <c...@email.com> wrote: >>> >>>> Hi JG, >>>> >>>> Speed is now down to 18 rows/sec/table per process. >>>> >>>> Here is a regionserver log that is serving two of the regions: >>>> >>>> http://pastebin.com/Hx5se0hz >>>> >>>> Here is the GC Log from the same server: >>>> >>>> http://pastebin.com/ChrRvxCx >>>> >>>> Here is the master log: >>>> >>>> http://pastebin.com/L1Kn66qU >>>> >>>> The thrift server logs have nothing in them in the same time period. >>>> >>>> Thanks in advance! >>>> >>>> -chris >>>> >>>> On Apr 28, 2010, at 7:32 PM, Jonathan Gray wrote: >>>> >>>>> Hey Chris, >>>>> >>>>> That's a really significant slowdown. I can't think of anything >> obvious that would cause that in your setup. >>>>> >>>>> Any chance of some regionserver and master logs from the time it was >> going slow? Is there any activity in the logs of the regionservers hosting >> the regions of the table being written to? >>>>> >>>>> JG >>>>> >>>>>> -----Original Message----- >>>>>> From: Christopher Tarnas [mailto:c...@tarnas.org] On Behalf Of Chris >>>>>> Tarnas >>>>>> Sent: Wednesday, April 28, 2010 6:27 PM >>>>>> To: hbase-user@hadoop.apache.org >>>>>> Subject: EC2 + Thrift inserts >>>>>> >>>>>> Hello all, >>>>>> >>>>>> First, thanks to all the HBase developers for producing this, it's a >>>>>> great project and I'm glad to be able to use it. >>>>>> >>>>>> I'm looking for some help and hints here with insert performance help. >>>>>> I'm doing some benchmarking, testing how I can scale up using HBase, >>>>>> not really looking at raw speed. The testing is happening on EC2, >> using >>>>>> Andrew's scripts (thanks - those were very helpful) to set them up and >>>>>> with a slightly customized version of the default AMIs (added my >>>>>> application modules). I'm using HBase 20.3 and Hadoop 20.1. I've >> looked >>>>>> at the tips in the Wiki and it looks like Andrew's scripts are already >>>>>> setup that way. >>>>>> >>>>>> I'm inserting into HBase from a hadoop streaming job that runs perl >> and >>>>>> uses the thrift gateway. I'm also using the Transactional tables so >>>>>> that alone could be the case, but from what I can tell I don't think >>>>>> so. LZO compression is also enabled for the column families (much of >>>>>> the data is highly compressible). My cluster has 7 nodes, 5 >>>>>> regionservers, 1 master and 1 zookeeper. The regionservers and master >>>>>> are c1.xlarges. Each regionserver has the tasktrackers that runs the >>>>>> hadoop streaming jobs, and regionserver also runs its own thrift >>>>>> server. Each mapper that does the load talks to the localhost's thrift >>>>>> server. >>>>>> >>>>>> The Row keys a fixed string + an incremental number then the order of >>>>>> the bytes are reversed, so runA123 becomes 321Anur. I though of using >>>>>> murmur hash but was worried about collisions. >>>>>> >>>>>> As I add more insert jobs, each jobs throughput goes down. Way down. I >>>>>> went from about 200 row/sec/table per job with one job to about 24 >>>>>> rows/sec/table per job with 25 running jobs. The servers are mostly >>>>>> idle. I'm loading into two tables, one has several indexes and I'm >>>>>> loading into three column families, the other has no indexes and one >>>>>> column family. Both tables only currently have two region each. >>>>>> >>>>>> The regionserver that serves the indexed table's regions is using the >>>>>> most CPU but is 87% idle. The other servers are all at ~90% idle. >> There >>>>>> is no IO wait. the perl processes are barely ticking over. Java on the >>>>>> most "loaded" server is using about 50-60% of one CPU. >>>>>> >>>>>> Normally when I do load in a pseudo-distrbuted hbase (my development >>>>>> platform) perl's speed is the limiting factor and uses about 85% of a >>>>>> CPU. In this cluster they are using only 5-10% of a CPU as they are >> all >>>>>> waiting on thrift (hbase). When I run only 1 process on the cluster, >>>>>> perl uses much more of a CPU, maybe 70%. >>>>>> >>>>>> Any tips or help in getting the speed/scalability up would be great. >>>>>> Please let me know if you need any other info. >>>>>> >>>>>> As I send this - it looks like the main table has split again and is >>>>>> being served by three regionservers.. My performance is going up a bit >>>>>> (now 35 rows/sec/table per processes), but still seems like I'm not >>>>>> using the full potential of even the limited EC2 system, no IO wait >> and >>>>>> lots of idle CPU. >>>>>> >>>>>> >>>>>> many thanks >>>>>> -chris >>>>>> >>>>>> >>>>>> >>>>>> >>>>>> >>>>> >>>> >> >> > > > -- > Todd Lipcon > Software Engineer, Cloudera
