[zeromq-dev] Java/Windows issues

[Martin Sustrik]

Steven McCoy wrote:
> 2009/8/21 Martin Sustrik <sustrik at fastmq.com>:
>> True. However, you should take performance impact of such architecture
>> into account. With multicast, each box on the network has to process all
>> messages and drop those that are not needed. Filtering is done on
>> software level and is often expensive - say regexp evaluation.
> 
> As demonstrated with Rendezvous v8 you can implement a communication
> channel that incorporates multiple wire transports (TRDP or PGM in
> this instance) that implement segmentation of the higher level
> namespace.  This allows a very basic level of hardware offload,
> although often hindered by OS and hardware limitations.

Exactly. I've spoke to people in couple of major networking HW companies 
about possible increase of offload capabilities directly in networking 
hardware for this kind of scenarios (like routing/filtering of multicast 
packets in switches/NICs based on ranges of multicast addresses or 
bitmasks, which in turn can be mapped to a subset of regexp use cases). 
Still, it would need much market pressure to get it done. At the moment 
we'll have to do with basic HW routing/filtering + software-level regexp 
processing.

> On a side note, TIBCO MDS uses hashes for high speed subject
> evaluation, with a hashing algorithm optimized for the MDS namespace.

Yup. Same plan here. Here's the design document:

http://www.zeromq.org/whitepapers:message-matching

Martin