Raghu Angadi 写道: > We don't expect such drastic slow down.. if b/w you report is the over > all throughput. > > What are the exact command line you ran? Also how is b/w you report > measured? > > the command as below: bin/hadoop jar hadoop-0.16.5-dev-test.jar TestDFSIO -write -nrFiles n -fileSize 2048 -bufferSize 65536 -resFile resfile
*TestDFSIO* (/src/test/org/apache/hadoop/fs/TestDFSIO.java) > Raghu. > > Samuel Guo wrote: > >> Hi all, >> >> I run the *TestDFSIO* benchmark on a simple cluster of 2 nodes. >> The file size is the same in all cases 2GB. >> The number of files tried is 1,2,4,8(only write). >> The bufferSize is 65536 bytes. >> The file replication is 1. >> >> the results as below: >> >> files 1 2 4 8 >> >> write -- Throughout(mb/s) 52.89 52.31 23.06 22.25 >> -- Avg IO rate(mb/s) 54.18 53.23 24.03 22.77 >> >> >> read -- Throughout(mb/s) 79.17 60.77 20.15 >> -- Avg IO rate(mb/s) 79.18 61.33 22.01 >> >> >> It doesn't seem good. When the number of writes is larger than the size >> of the cluster, everything become worse. >> >> Can anyone explain why everything is getting very slow when the the >> number of writes is close to or >> larger than the size of the cluster? >> >> Is there something wrong with my test or the cluster settings? >> >> Hope for your reply. >> >> regards, >> >> Samuel >> regards, Samuel
