[
https://issues.apache.org/jira/browse/IGNITE-3060?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15270183#comment-15270183
]
Semen Boikov commented on IGNITE-3060:
--------------------------------------
Merged into master several optimizations: reduced history size, got rid of
unnecessary message copies, removed buffered output streams, ignore duplicated
messages, do not process client heartbeats in RingMessageWorker (commit
f0daea9). With these changes was able to connect 500 clients to server with
-Xmx4g and only ~1g of memory was used (before got OOM with 300 clients).
Now need to do more realistic test with more servers/client to understand
client-per-server limit.
> Discovery: optimize resource usage for client connections
> ---------------------------------------------------------
>
> Key: IGNITE-3060
> URL: https://issues.apache.org/jira/browse/IGNITE-3060
> Project: Ignite
> Issue Type: Task
> Components: general
> Reporter: Semen Boikov
> Assignee: Semen Boikov
> Priority: Critical
> Fix For: 1.6
>
>
> Tried to run single server + 300 clients nodes, identified some issues which
> can be easily fixed:
> - default size of messages history which is used to support graceful clients
> reconenct is too large (10 * 1024)
> - now there is queue per client connection, when message is sent to clients
> copy is created for each queue, only one copy can be created
> - each client message worker has buffered output stream, now buffer size if
> pretty large, actually looks like this buffered output stream is not really
> needed
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
