I have a twisted Perspective Broker based server. Recently with an increased number of clients we are seeing a decrease in throughput. Profiling the server it seems that it is spending the majority of time sending messages to the clients (unsuprisingly).
PyPy might speed this up quite a bit; http://speed.pypy.org suggests PyPy can send PB messages 3 times as fast as CPython. You could also switch to a protocol that has a faster serialization, which shouldn't be too hard if your usage of PB is sufficiently simple.
Messages are sent to the client from the server using a callRemote, invoked from the event queue using reactor.callFromThread - nice and simple. What designs have people used to improve throughput in a heavily loaded server? My experience is telling me that I need to implement some sort of thread/process pool to divide the messaging to clients. Various documents on the twisted website strongly suggets that this is not a good idea.
Threading won't help much if you're CPU bound, since Python can only run one Python thread at once. Process pooling probably would help, though will involve a more complex infrastructure. Which part of the docs made you think that process pooling is a bad idea?