BTW, when testing this branch, we found some performance issues about HDFS Client: 1. we reduced the DFS client's heap allocation from 45% to 27% in HDFS-14535 [1]; 2. we also increased get throughput by 17.8% in disabled block cache case in HDFS-14541[2]. In theory, it should also helps a lot (especially p99/p999) even if RS has a high cacheHitRatio.
I think the next HDFS 2.8 release will include those patches, they're very good points for our HBase performance. [1]. https://issues.apache.org/jira/browse/HDFS-14535 [2]. https://issues.apache.org/jira/browse/HDFS-14541?focusedCommentId=16866472&page=com.atlassian.jira.plugin.system.issuetabpanels%3Acomment-tabpanel#comment-16866472 Thanks. On Tue, Jun 18, 2019 at 12:05 PM OpenInx <open...@gmail.com> wrote: > The HBASE-21879 has lots of changes: 123 files changed, 5833 > insertions(+), 3015 deletions(-). > Currently we developed this issue based on master branch, and expect to > release it in future HBase3.x. > Of course, if branch-2 want this feature we can do the backport, should > have some conflicts now but I > don't think it would be hard to fix because I believe the branch-2 > shouldn't have so much diff with master now > (at least in read path). > The first priority thing for now, I think it would be merging the > HBASE-21879 branch to master branch > before diverging. After that, I can do the backport. > > Thanks for your suggestion, Guanghao ! > > > > On Tue, Jun 18, 2019 at 11:39 AM Guanghao Zhang <zghao...@gmail.com> > wrote: > >> This is a improvement not a new feature? So backport to branch-2, too? >> >> OpenInx <open...@gmail.com> 于2019年6月17日周一 下午2:45写道: >> >> > Dear HBase dev: >> > >> > In HBASE-21879[1], we redesigned the offheap read path: read the >> HFileBlock >> > from HDFS to pooled offheap >> > ByteBuffers directly, while before HBASE-21879 we just read the >> HFileBlock >> > to heap which would still lead >> > to high GC pressure. >> > >> > After few months of development and testing, all subtasks have been >> > resovled now except the HBASE-21946[2] >> > (It depends on HDFS-14483[3] and our HDFS teams are working on this, we >> > expect the HDFS-14483 to be included >> > in hadoop 2.8.6 and after that the HBASE-21946 will get resolved). we >> think >> > the feature is stable enough now and it's >> > time to merge branch HBASE-21879 back to master now. >> > >> > We have designed 3 test cases to prove the performance improvment with >> > HBASE-21879: >> > 1. Disabled BlockCache, which means the cacheHitRatio is 0%; >> > 2. CacheHitRatio~65%; >> > 3. CachehHitRatio~100%; >> > >> > In our performance results[4], we can see that: the case#1 have an great >> > performance improvement >> > ( >> > *throughput increased about 17%, heap allocation decreased about 95%, >> Young >> > generaion size decreased about 81.7%*), that's because after HBASE-21879 >> > all reads will allocate from pooled offheap bytebuffers >> > and almost no heap allocation, while before HBASE-21879 the read path >> will >> > create so many heap allocations. >> > On the other hand, from the testing results of case#2 and case#3 we can >> > also see that: >> > >> > *As the cacheHitRatioincreasing, the difference between >> before-HBASE-21879 >> > and after-HBASE-21879 will decrease, when cacheHitRatio is 100%, they >> > almost have no much difference in both throughput and latency.* >> > >> > For more details please see the document[4]. Thanks >> > Anoop/Ram/DuoZhang/Stack/GuanghaoZhang very much >> > for your meticulous work (Suggession, discussion, patch reviewing, doc >> > reviewing etc). >> > >> > Please vote >> > >> > [] +1 >> > [] +0/-0 >> > [] -1 Do not merge the branch back because... >> > >> > Thanks. Any suggestions are welcomed. >> > >> > [1] https://issues.apache.org/jira/browse/HBASE-21879 >> > [2] https://issues.apache.org/jira/browse/HBASE-21946 >> > [3] https://issues.apache.org/jira/browse/HDFS-14483 >> > [4] >> > >> > >> https://docs.google.com/document/d/1xSy9axGxafoH-Qc17zbD2Bd--rWjjI00xTWQZ8ZwI_E >> > >> >