[
https://issues.apache.org/jira/browse/CASSANDRA-4277?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13281889#comment-13281889
]
Brandon Williams commented on CASSANDRA-4277:
---------------------------------------------
bq. 512 is really really high and the context switch will be really high and
you will not be seeing a lot of advantage with this.
I agree. Logically the optimal amount is the number of cores, but benchmarking
revealed better results with oversubscription, hence it defaulting to cores*4
now.
> hsha default thread limits make no sense, and yaml comments look confused
> -------------------------------------------------------------------------
>
> Key: CASSANDRA-4277
> URL: https://issues.apache.org/jira/browse/CASSANDRA-4277
> Project: Cassandra
> Issue Type: Bug
> Components: Core
> Reporter: Peter Schuller
>
> The cassandra.yaml states with respect to {{rpc_max_threads}}:
> {code}
> # For the Hsha server, the min and max both default to quadruple the number of
> # CPU cores.
> {code}
> The code seems to indeed do this. But this makes, as far as I can tell, no
> sense what-so-ever since the number of concurrent RPC threads you need is a
> function of the throughput and the average latency of requests (that includes
> synchronously waiting on network traffic).
> Defaulting to anything having to do with CPU cores seems inherently wrong. If
> a default is non-static, a closer guess might be to look at thread stack size
> and heap size and infer what "might" be reasonable.
> *NOTE*: The effect of having this too low, is "strange" (if you don't know
> what's going on) latencies observed form the client on all thrift requests
> (*any* thrift request, including e.g. {{describe_ring()}}), that isn't
> visible in any latency metric exposed by Cassandra. This is why I consider
> this "major", since unwitting users may be seeing detrimental performance for
> no good reason.
> In addition, I read this about async:
> {code}
> # async -> Nonblocking server implementation with one thread to serve
> # rpc connections. This is not recommended for high throughput use
> # cases. Async has been tested to be about 50% slower than sync
> # or hsha and is deprecated: it will be removed in the next major
> release.
> {code}
> This makes even less sense. Running with *one* rpc thread limits you to a
> single concurrent request. How was that 50% number even attained? By
> single-node testing being completely CPU bound locally on a node? The actual
> effect should be "stupidly slow" in any real situation with lots of requests
> on a cluster of many nodes and network traffic (though I didn't test that) -
> especially in the event of any kind of hiccup like a node doing GC. I agree
> that if the above is true, async should *definitely* be deprecated, but the
> reasons seem *much* stronger than implied.
> I may be missing something here, in which case I apologize,, but I
> specifically double-checked after I fixed this setting on on our our clusters
> after seeing exactly the expected side-effect of having it be too low. I
> always was under the impression that rpc_max_threads affects the number of
> RPC requests running concurrently, and code inspection (it being used for the
> worker thread limit) + the effects of client-observed latency is consistent
> with my understanding.
> I suspect the setting was set strangely by someone because the phrasing of
> the comments in {{cassandra.yaml}} strongly suggest that this should be tied
> to CPU cores, hiding the fact that this really has to do with the number of
> requests that can be serviced concurrently regardless of implementation
> details of thrift/networking being sync/async/etc.
--
This message is automatically generated by JIRA.
If you think it was sent incorrectly, please contact your JIRA administrators:
https://issues.apache.org/jira/secure/ContactAdministrators!default.jspa
For more information on JIRA, see: http://www.atlassian.com/software/jira
