[
https://issues.apache.org/jira/browse/CAMEL-19060?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Otavio Rodolfo Piske updated CAMEL-19060:
-----------------------------------------
Description:
When using the Seda producer/consumer on a large core count machine, it seems
that Camel's throughput peak at around 9 producer / 9 consumer threads and
after that starts to decrease [1]. This seems to indicate some kind of
coherence issue - the cost to keep the threads coherent exceeds the system's
ability to handle the load / use the system's free capacity.
1. This, of course, taking in consideration that the thread pool had been
adjusted and the machine has enough capacity. Both of which were true.
The attachments contains `perf` reports of the analysis of Camel 3.18.5, but
similar patterns also happen w/ 4.0.0-M1. In particular, note the low
"Instruction per cycle" reported on the attachment [^perf-stats.txt] and the
high incidence of Last Level Cache (LLC) misses to Remote Cache (HITM) in the
attachment [^stats.txt].
was:
When using the Seda producer/consumer on a large core count machine, it seems
that Camel's throughput peak at around 9 producer / 9 consumer threads and
after that starts to decrease [1]. This seems to indicate some kind of
coherence issue - the cost to keep the threads coherent exceeds the system's
ability to handle the load / use the system's free capacity.
1. This, of course, taking in consideration that the thread pool had been
adjusted and the machine has enough capacity. Both of which were true.
> Scalability ceiling
> -------------------
>
> Key: CAMEL-19060
> URL: https://issues.apache.org/jira/browse/CAMEL-19060
> Project: Camel
> Issue Type: Task
> Affects Versions: 3.18.5, 4.0-M1
> Reporter: Otavio Rodolfo Piske
> Assignee: Otavio Rodolfo Piske
> Priority: Major
> Attachments: perf-stats.txt, stats.txt
>
>
> When using the Seda producer/consumer on a large core count machine, it seems
> that Camel's throughput peak at around 9 producer / 9 consumer threads and
> after that starts to decrease [1]. This seems to indicate some kind of
> coherence issue - the cost to keep the threads coherent exceeds the system's
> ability to handle the load / use the system's free capacity.
>
> 1. This, of course, taking in consideration that the thread pool had been
> adjusted and the machine has enough capacity. Both of which were true.
> The attachments contains `perf` reports of the analysis of Camel 3.18.5, but
> similar patterns also happen w/ 4.0.0-M1. In particular, note the low
> "Instruction per cycle" reported on the attachment [^perf-stats.txt] and the
> high incidence of Last Level Cache (LLC) misses to Remote Cache (HITM) in the
> attachment [^stats.txt].
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
