[
https://issues.apache.org/jira/browse/HBASE-26067?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17448499#comment-17448499
]
Wellington Chevreuil commented on HBASE-26067:
----------------------------------------------
{quote}Do you have any initial perf numbers on this? I'm curious how it
compares to Stephen's original results.{quote}
We have done some ycsb runs with write only workloads and compared the ycsb
throughput/latency metrics between runs with and without this feature. When
enabling the feature, we also disabled HBoss) Here's an example:
SFT disabled/HBoss enabled :
{noformat}
[OVERALL], RunTime(ms), 2556932
[OVERALL], Throughput(ops/sec), 3910.9370135772087
[TOTAL_GCS_PS_Scavenge], Count, 3660
[TOTAL_GC_TIME_PS_Scavenge], Time(ms), 8642
[TOTAL_GC_TIME_%_PS_Scavenge], Time(%), 0.3379831767133424
[TOTAL_GCS_PS_MarkSweep], Count, 1
[TOTAL_GC_TIME_PS_MarkSweep], Time(ms), 23
[TOTAL_GC_TIME_%_PS_MarkSweep], Time(%), 8.995155131227581E-4
[TOTAL_GCs], Count, 3661
[TOTAL_GC_TIME], Time(ms), 8665
[TOTAL_GC_TIME_%], Time(%), 0.33888269222646517
[CLEANUP], Operations, 20
[CLEANUP], AverageLatency(us), 86.9
[CLEANUP], MinLatency(us), 2
[CLEANUP], MaxLatency(us), 1337
[CLEANUP], 95thPercentileLatency(us), 236
[CLEANUP], 99thPercentileLatency(us), 1337
[INSERT], Operations, 10000000
[INSERT], AverageLatency(us), 2552.2381777
[INSERT], MinLatency(us), 1146
[INSERT], MaxLatency(us), 1753087
[INSERT], 95thPercentileLatency(us), 3025
[INSERT], 99thPercentileLatency(us), 20623
[INSERT], Return=OK, 10000000 {noformat}
SFT enablde/HBoss Disabled:
{noformat}
[OVERALL], RunTime(ms), 2420289
[OVERALL], Throughput(ops/sec), 4131.737986661923
[TOTAL_GCS_PS_Scavenge], Count, 3639
[TOTAL_GC_TIME_PS_Scavenge], Time(ms), 8213
[TOTAL_GC_TIME_%_PS_Scavenge], Time(%), 0.33933964084454377
[TOTAL_GCS_PS_MarkSweep], Count, 1
[TOTAL_GC_TIME_PS_MarkSweep], Time(ms), 23
[TOTAL_GC_TIME_%_PS_MarkSweep], Time(%), 9.502997369322424E-4
[TOTAL_GCs], Count, 3640
[TOTAL_GC_TIME], Time(ms), 8236
[TOTAL_GC_TIME_%], Time(%), 0.34028994058147605
[CLEANUP], Operations, 20
[CLEANUP], AverageLatency(us), 74.85
[CLEANUP], MinLatency(us), 1
[CLEANUP], MaxLatency(us), 1047
[CLEANUP], 95thPercentileLatency(us), 261
[CLEANUP], 99thPercentileLatency(us), 1047
[INSERT], Operations, 10000000
[INSERT], AverageLatency(us), 2415.8930098
[INSERT], MinLatency(us), 1129
[INSERT], MaxLatency(us), 1608703
[INSERT], 95thPercentileLatency(us), 2979
[INSERT], 99thPercentileLatency(us), 19231
[INSERT], Return=OK, 10000000 {noformat}
This, of course, is what's observed from the ycsb client. We had not yet
measured performance/throughput of individual internal operations that are
probably most impacted, such as compactions, splits/merges, snapshots, to name
a few.
> Change the way on how we track store file list
> ----------------------------------------------
>
> Key: HBASE-26067
> URL: https://issues.apache.org/jira/browse/HBASE-26067
> Project: HBase
> Issue Type: Umbrella
> Components: HFile
> Reporter: Duo Zhang
> Assignee: Duo Zhang
> Priority: Major
>
> Open a separated jira to track the work since it can not be fully included in
> HBASE-24749.
> I think this could be a landed prior to HBASE-24749, as if this works, we
> could have different implementations for tracking store file list.
--
This message was sent by Atlassian Jira
(v8.20.1#820001)
