[
https://issues.apache.org/jira/browse/HBASE-21657?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16735455#comment-16735455
]
Zheng Hu commented on HBASE-21657:
----------------------------------
Those days, I made some tests for the above cases:
||HBaseVersion||Storage||QPS&Latency||FlameGraph||Comment||
|HBase2.0.4|SSD|[^HBase2.0.4-ssd-10000000-rows-qps-latency.png]|[^HBase2.0.4-ssd-10000000-rows-flamegraph.svg]|regionCount=100,
rows=10^7, dataSizeOfTable=1.5GB, cacheHitRatio=100%|
|HBase2.0.4 +
patch.v2|SSD|[^HBase2.0.4-patch-v2-ssd-10000000-rows-qps-and-latency.png]|[^HBase2.0.4-patch-v2-ssd-10000000-rows.svg]|regionCount=100,
rows=10^7, dataSizeOfTable=1.5GB, cacheHitRatio=100%|
|HBase1.4.9|SSD|[^HBase1.4.9-ssd-10000000-rows-qps-latency.png]|[^HBase1.4.9-ssd-10000000-rows-flamegraph.svg]|regionCount=100,
rows=10^7, dataSizeOfTable=1.5GB, cacheHitRatio=100%|
Besides, I made an overview stastics :
!overview-statstics-1.png!
we can see that: *the performance of hbase1.4.9 is almost the same as the
HBase2.0.4 with patch.v2.* So, I think we are in the right direction to
optimize hbase2.0 performance.
Now the problem is: how do we write the patch ? IMO, we can move the
getSerializedSize() (without no tag param) and heapSize() to the Cell
interface for eliminating the instanceof and class casting, also the
predetermined size of results arraylist will help a lot, not necessary to be
1000, we can choose the Min(rows, 512) for avoiding cost too much memory for a
scan with huge rows.
Haven't looked at the inline method in detail yet, will try to do this.
[~stack] FYI
> PrivateCellUtil#estimatedSerializedSizeOf has been the bottleneck in 100%
> scan case.
> ------------------------------------------------------------------------------------
>
> Key: HBASE-21657
> URL: https://issues.apache.org/jira/browse/HBASE-21657
> Project: HBase
> Issue Type: Bug
> Components: Performance
> Reporter: Zheng Hu
> Assignee: Zheng Hu
> Priority: Major
> Fix For: 3.0.0, 2.2.0, 2.1.3, 2.0.5
>
> Attachments: HBASE-21657.v1.patch, HBASE-21657.v2.patch,
> HBase1.4.9-ssd-10000000-rows-flamegraph.svg,
> HBase1.4.9-ssd-10000000-rows-qps-latency.png,
> HBase2.0.4-patch-v2-ssd-10000000-rows-qps-and-latency.png,
> HBase2.0.4-patch-v2-ssd-10000000-rows.svg,
> HBase2.0.4-ssd-10000000-rows-flamegraph.svg,
> HBase2.0.4-ssd-10000000-rows-qps-latency.png, HBase2.0.4-with-patch.v2.png,
> HBase2.0.4-without-patch-v2.png, hbase2.0.4-ssd-scan-traces.2.svg,
> hbase2.0.4-ssd-scan-traces.svg, hbase20-ssd-100-scan-traces.svg,
> overview-statstics-1.png
>
>
> We are evaluating the performance of branch-2, and find that the throughput
> of scan in SSD cluster is almost the same as HDD cluster. so I made a
> FlameGraph on RS, and found that the
> PrivateCellUtil#estimatedSerializedSizeOf cost about 29% cpu, Obviously, it
> has been the bottleneck in 100% scan case.
> See the [^hbase20-ssd-100-scan-traces.svg]
> BTW, in our XiaoMi branch, we introduce a
> HRegion#updateReadRequestsByCapacityUnitPerSecond to sum up the size of cells
> (for metric monitor), so it seems the performance loss was amplified.
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
