[zeromq-dev] Yahoo shows big performance improvement by switching Storm from ZeroMQ to Netty

[Pieter Hintjens]

Their test case is so specific it's hard to draw general conclusions.
They're using acking, so ZeroMQ isn't doing any batching. We know that
ZeroMQ has a fairly high cost on individual messages, which pays off
when you stream data.

What ZeroMQ patterns are they using? What are the message sizes? How
many messages are being sent? Do we have any sample code that can be
measured? What version of ZeroMQ are they using? Is it by any chance
2.x, with its old inefficient pub-sub model?

And, "the new default settings allow Netty to run twice as fast as the
default setting for zeromq" suggests that a tuned Netty runs faster in
this use case than an untuned ZeroMQ.

it's not convincing science.

Now, if someone would extract that use case and show reproducible
figures, that would be interesting.

-Pieter

On Mon, Apr 7, 2014 at 12:57 PM, Steven McCoy <steven.mccoy at miru.hk> wrote:
> On 7 April 2014 04:10, Asquith, Edward <edward.asquith at baesystems.com>
> wrote:
>>
>> Obviously Netty supports a narrower set of use cases, but being pure Java
>> this certainly caught my attention; better performance and a simplified
>> deployment model is a compelling combo.
>>
>>
>
> I think JeroMQ would have done the same and more, JNI is really slow.
>
> --
> Steve-o
>
> _______________________________________________
> zeromq-dev mailing list
> zeromq-dev at lists.zeromq.org
> http://lists.zeromq.org/mailman/listinfo/zeromq-dev
>