On 04/02/2010 06:09 PM, Chen Bangzhong wrote: > my switch is Dell 2724. > > I'm not a network admin, and I don't have the ability to know how congested your network is from that(nor do I think it is possible since there's going to be a lot of other factors).
Try running the test on a single machine using the miniCluster flag, this should eliminate network transfer as an issue. If despite the fact you're running everything on a single machine you get a high throughput on that your network is likely the issue. If on the other hand throughput goes down significantly the problem lies elsewhere. > --在 2010年4月2日 下午5:04，Chen Bangzhong <bangzh...@gmail.com>写道： > > >> >> 在 2010年4月2日 下午4:58，Juhani Connolly <juh...@ninja.co.jp>写道： >> >> You're results seem very low, but your system specs are also quite >> >>> moderate. >>> >>> On 04/02/2010 04:46 PM, Chen Bangzhong wrote: >>> >>>> Hi, All >>>> >>>> I am benchmarking hbase. My HDFS clusters includes 4 servers (Dell 860, >>>> >>> with >>> >>>> 2 GB RAM). One NameNode, one JobTracker, 2 DataNodes. >>>> >>>> My HBase Cluster also comprise 4 servers too. One Master, 2 region and >>>> >>> one >>> >>>> ZooKeeper. (Dell 860, with 2 GB RAM) >>>> >>>> >>> While I'm far from being an authority on the matter, running >>> datanodes+regionservers together should help performance >>> Try making your 2 datanodes + 2 regionservers into 4 servers running >>> both data/region. >>> >>> >> I will try to run datanode and region server on the same server. >> >> >> >>>> I runned the org.apache.hadoop.PerformanceEvaluation on the ZooKeeper >>>> server. the ROW_LENGTH was changed from 1000 to ROW_LENGTH = 100*1024; >>>> So each value will be 100k in size. >>>> >>>> hadoop version is 0.20.2, hbase version is 0.20.3. dfs.replication set >>>> >>> to 1. >>> >>>> >>> Setting replication to 1 isn't going to give results that are very >>> indicative of a "real" application, making it questionable as a >>> benchmark. If you intend to run on a single replica at release, you'll >>> be at high risk of data loss. >>> >>> >> Since I have only 2 data nodes, I set replication to 1. In production, it >> will be set to 3. >> >> >> >>>> The following is the command line: >>>> >>>> bin/hbase org.apache.hadoop.hbase.PerformanceEvaluation --nomapred >>>> --rows=10000 randomWrite 20. >>>> >>>> It tooks about one hour to complete the test(3468628 ms), about 60 >>>> >>> writes >>> >>>> per second. It seems the performance is disappointing. >>>> >>>> Is there anything I can do to make hbase perform better under 100k size >>>> >>> ？I >>> >>>> didn't try the method mentioned in the performance wiki yet, because I >>>> thought 60writes/sec is too low. >>>> >>>> >>>> >>> Do you mean *over* 100k size? >>> 2GB ram is pretty low and you'd likely get significantly better >>> performance with it, though on this scale it probably isn't a >>> significant problem. >>> >>> >> the data size is exactly 100k size. >> >> >> >>>> If the value size is 1k, hbase performs much better. 200000 >>>> >>> sequencewrite >>> >>>> tooks about 16 seconds, about 12500 writes/per second. >>>> >>>> >>>> >>> Comparing sequencewrite performance with randomwrite isn't a helpful >>> indicator. Do you have randomWrite results for 1k values? The way your >>> performance degrades with the size of the records seems like you may >>> have a bottleneck at network transfer? What's rack locality like and how >>> much bandwidth do you have between the servers? >>> >>>> Now I am trying to benchmark using two clients on 2 servers, no result >>>> >>> yet. >>> >>>> >>>> >>> >> for 1k datasize, the sequencewrite performance and randomWrite performance >> is about the same. All my servers are under one switch, don't know the >> switch bandwidth yet. >> >> >> >>> You're already running 20 clients on your first server with the >>> PerformanceEvaluation. Do you mean you intend to run 20 on each? >>> >>> >> In fact, it is 20 threads on one machine. >> >> >>> Hopefully someone with better knowledge can give a better answer but my >>> guess is that you have a network transfer transfer. Try doing further >>> tests with randomWrite and decreasing value sizes and see if the time >>> correlates to the total amount of data written. >>> >>> >>> >> >
