I ran PE on a i7, 12GB, 2x1TB machine in a fully-distributed 1 node setup. I restarted HBase between each test.
sequentialWrite 1: 12306 rows/sec sequentialRead 1: 3485 (I do the randomRead test before the randomWrite, because it's not "perfectly" random as some values will be read or wrote more than once and others won't be touched at all) randomRead 1: 696 scan 1: 36839 randomeWrite: 1: 13095 This is what I also expect from these tests. J-D 2010/6/23 史英杰 <[email protected]>: > Hi, > Sequential write is about 7000row/s sequential Read 20000 rows/s > random write is about 8000 row/s random read 17000 rows/s > So how about your numbers? > 2010/6/24 Jean-Daniel Cryans <[email protected]> > >> When I run PE, random/seqRead are always much slower than >> random/seqWrite so it seems to be something about your env or you are >> confused by the output of that test. >> >> What are the numbers you see? >> >> J-D >> >> On Wed, Jun 23, 2010 at 5:32 AM, 史英杰 <[email protected]> wrote: >> > Hi, All >> > Recently I did some tests on HBase, using the performance eveluation >> > package in HBase 0.20.3. There are some situations I can't understand. I >> > found that read perform better than write, no matter sequential read or >> > random read during the test. But in BigTable's paper, writes perform >> better >> > than reads, because each tablet server appends all incoming writes to a >> > single commit log and uses group commit to stream these writes >> efficiently >> > to GFS. In HBase, we set the autoFlush=false, and the flush size is 64M, >> so >> > writes should perform better than reads, but the result is just the >> > opposite. Please help me to explain this phenomenon, thanks a lot! >> > >> > Yingjie >> > >> >
