cassandra-commits mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Thom Valley (JIRA)" <j...@apache.org>
Subject [jira] [Commented] (CASSANDRA-12104) Handle coalesce efforts for inter-dc traffic discretely from intra-dc traffic
Date Tue, 05 Jul 2016 21:15:11 GMT

    [ https://issues.apache.org/jira/browse/CASSANDRA-12104?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15363265#comment-15363265
] 

Thom Valley commented on CASSANDRA-12104:
-----------------------------------------

Ariel,

Something to consider in that is both what Jeremiah stated above, but also the fact that DC
location can heavily impact both available bandwidth and latency.  

In the example I have, there are two DCs in the US, relatively close to each other (30ms apart)
with almost 10G bandwidth between them, but the remote DCs vary from 70ms up to over 200 ms
with substantially less bandwidth available.

The ideal scenario for most inter-dc throttles would be to control it by connection (by that
I mean each distinct inter-dc path).  But perhaps that is too complicated.

I will get the details from the customer on exactly what levels showed the benefits based
on which DC we were evaluating.


> Handle coalesce efforts for inter-dc traffic discretely from intra-dc traffic
> -----------------------------------------------------------------------------
>
>                 Key: CASSANDRA-12104
>                 URL: https://issues.apache.org/jira/browse/CASSANDRA-12104
>             Project: Cassandra
>          Issue Type: Improvement
>            Reporter: Thom Valley
>            Priority: Minor
>
> In relationship to CASSANDRA-8692, we have discovered that pushing coalescing windows
to the point where they have a positive impact on inter-dc traffic overhead appears to have
causes delays in intra-dc traffic (namely, quorum requests between nodes).  Having the same
coalescing strategy apply to all messages (especially intra-dc request/response messages)
seems like a bad idea.
> This was in a 5 DC environment with from 30 to 130 ms of latency between the DCs.  Local
network was entirely unrestricted 10G ethernet.
> Being able to apply different coalescing rules to those two classifications of traffic
would allow much more effective tuning of the coalescing strategies, save inter-dc bandwidth
while not having any impact on intra-dc message handling.



--
This message was sent by Atlassian JIRA
(v6.3.4#6332)

Mime
View raw message