[zeromq-dev] Multicore Magic

[Martin Sustrik]

Pieter,

> Presumably there is then a way to use shared memory to get full-speed
> message transfer between processes.  Martin, does that make sense?

No, there's not. There are multiple problems with shared memory:

1. Control events still have to be transferred via some IPC mechanism 
even if message payload is stored in shmem. So there's no way to improve 
latency to the level of inproc transport.

2. Allocating shmem is not free. Allocating shmem buffer for each 
message would turn to performance bottleneck.

3. It can make sense to allocate message payload in shmem for very large 
messages - so large that the cost of copying required by underlying IPC 
transport would exceed the cost of allocating shmem buffer.

In the latter case I would suggest allocating the shmem buffer manually 
and pass the pointer as a 0MQ message to the other peer.

Martin