Hi Dmitry. Looks like high network latency. Do you run this test with client and server on the same machine, or you test from another machine? May be over wireless?
2011/9/6 Дмитрий <[email protected]> > Hello everyone! > We started using hbase (hadoop) system and faced some performance issues. > Actually we are using hbase in pseudo distributed mode on one node. > We used Cloudera distribution pack of Hadoop on operating system CentOs 6 > with default configuration according to > https://ccp.cloudera.com/display/CDHDOC/HBase+Installation. > > So, we started to test them on random reading. > Test data contains one table. Each row has length about 10 Kb. Total data > size is 400 000 rows (or about 3,19 Gb). > Average random reading rate from one Thrift/Java API connection is 30 rows > per second, writing --- 250 rows per second. > If we use 4 connections, random reading rate increases to 120 rows per > second on each connection, or total 480 rows per second. > So, increasing connections involves increasing of random reading > performance on each connection. > However, standard tests from > http://wiki.apache.org/hadoop/Hbase/PerformanceEvaluation showed > performance 2000 rows per second for random reading. > Also we noticed, that overall node resources (io, cpu) are being used no > more than 3%. We have enough RAM (8G and 2 of them is free). > > Is there any rational explanation of this issue? > > Best regards, Dmitry Kangin. >
