[zeromq-dev] Thoughts on Linux End User Summit

[Martin Sustrik]

Vitaly Mayatskikh wrote:
> At Fri, 27 Nov 2009 10:08:21 +0100, Martin Sustrik wrote:
> 
>> Any thoughts on the issue anybody? Is there a section of messaging 
>> functionality that would benefit from being moved down to the kernel? 
>> Any experience with that kind of thing anybody?
> 
> Forwarder device can be implemented as a kernel module. It is evident,
> that mq with broker can benefit from moving code to kernel. It doesn't
> matter in brokerless systems, since both communication points lay in
> userspace.

The point the guy from BoA was making IMO, was not to move the whole 
middleware into the kernel. Rather he asked for minimising the number of 
user/kernel space transitions for networking-heavy applications.

One example of such optimisation would be to move the filtering code 
(see ZMQ_SUBSCRIBE option) to the kernel, so that incoming packets that 
nobody wants are dropped even before being moved to the user space.

Anyway, it would be worth of keeping track of how kernel community 
responds to the challenge.

> One of ways to speed up brokerless systems a bit is to use OFED.

There were attempts to use RDMA directly, however, the code never got to 
the mainline. Still, we've used SDP to run 0MQ on top of InfiniBand and 
10GbE. Have a look at the results here:

http://www.zeromq.org/results:ib-tests-v031
http://www.zeromq.org/results:10gbe-tests-v031

Martin