[jboss-jira] [JBoss JIRA] (JGRP-1717) Bundling reduces performance in scenario mimicking Inifinispan

[Bela Ban (JIRA)]

    [ https://issues.jboss.org/browse/JGRP-1717?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=12934690#comment-12934690 ] 

Bela Ban edited comment on JGRP-1717 at 1/9/14 9:13 AM:
--------------------------------------------------------

Tested perf with UPerf (measures throughput and latency)
Environment:
- JGroups master (3.5) with JGRP-1767 and JGRP-1768 included
- cluster-xx lab, 1Gbps network (~125MB/sec)
- 8 nodes: cluster01-08
- fast.xml (attached) with ip_ttl=1 *and queue for OOB thread pool*
- Each node invokes 20'000 RPCs: 80% sync GETs and 20% async PUTs
- numOwners=2

- Configs are:
|| A || B || C ||
| ignore_dont_bundle=false | ignore_dont_bundle=true | ignore_dont_bundle=true|
| {{+DONT_BUNDLE +OOB}} | {{+DONT_BUNDLE +OOB}} | {{-DONT_BUNDLE -OOB}} |



The results are (in avg. requests / sec / node):

|| Invoker threads || SSWT-A || SSWT-B || SSWT-C || TQ-A || TQ-B || TQ-C || SS-A || SS-B || SS-C ||
|   1 | 4754 | 4545 | 23 | *4809* | 4415 | 4647 | 4698 | *4718* | *4656* |
|   4 | 14699 | 14033 | 74 | *15122* | 13864 | *14079* | 15715 | *14275* | 13782 |
| 10 | 22948 | 21192 | 227 | *23232* | *22576* | *22312* | 22487 | 21920 | 18403 |
| 20 | 24431 | 23738 | 340 | 23988 | *29282* | *25522* | *25673* | 28228 | 20890 |

Comments:
* SSWT-B: messages are not bundled (even though ignore_don_bundle=true) as the code has an exception for this bundler
* SSWT-C: this bundler has a max_bundle_time of 30ms, so each request and response can have up to 30ms added to it, which is very bad for latency
** Running sync RPCs on SSWT without {{+DONT_BUNDLE | +OOB}} doesn't make any sense
* With option C, the cost of batching doesn't really show (although it *is* a bit slower than option B), as UPerf doesn't do anything real (unlike Infinispan).
** However, the slower perf of C against B is caused by all requests in a batch being processed *sequentially* and therefore responses sent before the next request is processed slow down overall speed
* Infinispan uses option B when invoking sync RPCs
                
      was (Author: belaban):
    Tested perf with UPerf (measures throughput and latency)
Environment:
- JGroups master (3.5) with JGRP-1767 and JGRP-1768 included
- cluster-xx lab, 1Gbps network (~125MB/sec)
- 8 nodes: cluster01-08
- fast.xml (attached) with ip_ttl=1 *and queue for OOB thread pool*
- Each node invokes 20'000 RPCs: 80% sync GETs and 20% async PUTs
- numOwners=2

- Configs are:
|| A || B || C ||
| ignore_dont_bundle=false | ignore_dont_bundle=true | ignore_dont_bundle=true|
| {{+DONT_BUNDLE +OOB}} | {{+DONT_BUNDLE +OOB}} | {{-DONT_BUNDLE -OOB}} |



The results are (in avg. requests / sec / node):

|| Invoker threads || SSWT-A || SSWT-B || SSWT-C || TQ-A || TQ-B || TQ-C || SS-A || SS-B || SS-C ||
|   1 | 4754 | 4545 | 23 | 4809 | 4415 | 4647 | 4698 | 4718 | 4656 |
|   4 | 14699 | 14033 | 74 | 15122 | 13864 | 14079 | 15715 | 14275 | 13782 |
| 10 | 22948 | 21192 | 227 | 23232 | 22576 | 22312 | 22487 | 21920 | 18403 |
| 20 | 24431 | 23738 | 340 | 23988 | 29282 | 25522 | 25673 | 28228 | 20890 |

Comments:
* SSWT-B: messages are not bundled (even though ignore_don_bundle=true) as the code has an exception for this bundler
* SSWT-C: this bundler has a max_bundle_time of 30ms, so each request and response can have up to 30ms added to it, which is very bad for latency
** Running sync RPCs on SSWT without {{+DONT_BUNDLE | +OOB}} doesn't make any sense
* With option C, the cost of batching doesn't really show (although it *is* a bit slower than option B), as UPerf doesn't do anything real (unlike Infinispan).
** However, the slower perf of C against B is caused by all requests in a batch being processed *sequentially* and therefore responses sent before the next request is processed slow down overall speed
* Infinispan uses option B when invoking sync RPCs
                  
> Bundling reduces performance in scenario mimicking Inifinispan
> --------------------------------------------------------------
>
>                 Key: JGRP-1717
>                 URL: https://issues.jboss.org/browse/JGRP-1717
>             Project: JGroups
>          Issue Type: Bug
>    Affects Versions: 3.3
>            Reporter: Radim Vansa
>            Assignee: Bela Ban
>             Fix For: 3.5
>
>         Attachments: benchmark-jgroups.xml, fast.xml, jgroups-udp.pdf, jgroups-udp.xml
>
>
> In Radargun benchmark of JGroups mimicking the behaviour of Infinispan we can see that performance with bundling on actually reduces the performance.
> See attached configurations and results.

--
This message is automatically generated by JIRA.
If you think it was sent incorrectly, please contact your JIRA administrators
For more information on JIRA, see: http://www.atlassian.com/software/jira