[zeromq-dev] Preliminary results of voting

Pieter Hintjens ph at imatix.com
Tue Sep 2 09:44:27 CEST 2008


On Tue, Sep 2, 2008 at 8:38 AM, Martin Sustrik <sustrik at fastmq.com> wrote:

> 1. Would it be enough to use wildcard only on the end of the selector
> string? (I.e. "animals.mammals.*" but not "animals.*.dogs".)

Presumably yes, at least in an initial version.

> 2. Would it be sufficient to have the topic hierarchy precompiled in
> advance and not to be able to add new topics on the fly?

The solution we used in OpenAMQ is to recompile topic bindings as they
hit new topics.  If you use the inverted bitmap technique, this means
a brand new topic causes a one-off latency spike as all bindings are
recompiled.  That can be largely fixed by cacheing the topic tree to
disk.  You'd still get the rare spike.  You could eliminate this by
not allowing new topics, so in effect getting your model of a
precompiled, fixed topic tree.

> The next desired item - surprisingly - is to lower the latency from
> current 30 us to 10 us (5 votes). I have no idea what conclusion to make
> from this. 30 us is as low as you can get on commodity hardware and
> standard networking stack.

This is the cost of pushing a single message through the stack and
back up the other side, right?  If you bundle messages you can cut
this latency.  Does 0MQ bundle outgoing messages, by emptying its
queue?  We're looking at doing this in OpenAMQ to improve throughput
(and cut latency).

-Pieter



More information about the zeromq-dev mailing list