[
https://issues.apache.org/jira/browse/HDFS-7358?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14206049#comment-14206049
]
stack commented on HDFS-7358:
-----------------------------
bq. So the synchronization is required.
What of Packet#writeTo did the buffer release?
bq. We already has close(), which is the public user API for closing the stream.
Ok.
Here's a few numbers:
With the feature turned off:
||threads||seconds||ops/second||
|10|133.662|7481.558|
|10|133.599|7485.086|
|10|134.046|7460.125|
|20|140.972|14187.215|
|20|141.949|14089.566|
|20|140.861|14198.395|
|100|153.941|64959.953|
|100|153.751|65040.223|
|100|153.372|65200.953|
With version of patch that does NOT synchronize Packet:
||threads||seconds||ops/second||
|10|126.219|7922.737|
|10|127.792|7825.216|
|10|124.829|8010.959|
|20|138.132|14478.904|
|20|137.051|14593.108|
|20|139.604|14326.236|
|100|149.311|66974.297|
|100|149.849|66733.844|
|100|149.537|66873.078|
Here is latest patch numbers:
||threads||seconds||ops/second||
|10|127.079|7869.121|
|10|128.357|7790.771|
|10|129.122|7744.614|
|20|135.525|14757.426|
|20|139.531|14333.731|
|20|135.595|14749.807|
|100|149.802|66754.781|
|100|149.262|66996.289|
|100|149.925|66700.016|
Threads in above are client threads. Actual number of writing and syncing
threads stays constant at 1 and 5. More threads just means more writing per
second.
Comparing the last run of 100 threads with the feature off vs the last run of
the latest patch I see more stalls, about same instructions per cycle but less
cycles so it comes out a bit better.
Perf summary on unpatched run:
{code}
Performance counter stats for '/home/stack/hbase/bin/hbase --config
/home/stack/conf_hbase
org.apache.hadoop.hbase.regionserver.wal.HLogPerformanceEvaluation -threads 100
-iterations 100000 -keySize 50 -valueSize 100':
587172.254075 task-clock # 3.666 CPUs utilized
18,700,961 context-switches # 0.032 M/sec
4,596,456 CPU-migrations # 0.008 M/sec
650,547 page-faults # 0.001 M/sec
891,035,644,874 cycles # 1.518 GHz
[83.31%]
674,789,502,548 stalled-cycles-frontend # 75.73% frontend cycles idle
[83.32%]
400,621,650,589 stalled-cycles-backend # 44.96% backend cycles idle
[66.74%]
422,912,592,386 instructions # 0.47 insns per cycle
# 1.60 stalled cycles per insn
[83.41%]
78,498,471,337 branches # 133.689 M/sec
[83.37%]
2,768,724,048 branch-misses # 3.53% of all branches
[83.26%]
160.168742742 seconds time elapsed
{code}
Here is patched version perf output.
{code}
Performance counter stats for '/home/stack/hbase/bin/hbase --config
/home/stack/conf_hbase
org.apache.hadoop.hbase.regionserver.wal.HLogPerformanceEvaluation -threads 100
-iterations 100000 -keySize 50 -valueSize 100':
556038.390042 task-clock # 3.550 CPUs utilized
18,699,748 context-switches # 0.034 M/sec
4,534,830 CPU-migrations # 0.008 M/sec
636,724 page-faults # 0.001 M/sec
843,860,285,154 cycles # 1.518 GHz
[83.29%]
642,851,753,015 stalled-cycles-frontend # 76.18% frontend cycles idle
[83.34%]
384,260,620,446 stalled-cycles-backend # 45.54% backend cycles idle
[66.66%]
392,462,867,299 instructions # 0.47 insns per cycle
# 1.64 stalled cycles per insn
[83.36%]
71,358,339,182 branches # 128.333 M/sec
[83.43%]
2,712,426,902 branch-misses # 3.80% of all branches
[83.29%]
156.646653202 seconds time elapsed
{code}
> Clients may get stuck waiting when using ByteArrayManager
> ---------------------------------------------------------
>
> Key: HDFS-7358
> URL: https://issues.apache.org/jira/browse/HDFS-7358
> Project: Hadoop HDFS
> Issue Type: Bug
> Components: hdfs-client
> Reporter: Tsz Wo Nicholas Sze
> Assignee: Tsz Wo Nicholas Sze
> Attachments: h7358_20141104.patch, h7358_20141104_wait_timeout.patch,
> h7358_20141105.patch, h7358_20141106.patch, h7358_20141107.patch,
> h7358_20141108.patch
>
>
> [~stack] reported that clients might get stuck waiting when using
> ByteArrayManager; see [his
> comments|https://issues.apache.org/jira/browse/HDFS-7276?focusedCommentId=14197036&page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel#comment-14197036].
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
