[erlang-questions] Re: 20k messages in 4s but want to go faster!

[Joel Reymont]

On Jul 13, 2009, at 8:58 AM, Jim Morris wrote:

> I think it is actually faster.

Have you measured it?

> My server uses a gen_fsm for each
> socket/client connection. When I get a message that needs  
> broadcasting/
> unicasting to all the other clients, I get a list of all the PIDS of
> the gen_fsm's, and send an erlang message to each of the pids,

That's what I (used to) do as well.


> when the 8,000 or so gen_fsm processes get that message they do a  
> tcp send
> to their client. That is getting the fastest throughput for me.

How do you know it's getting the fastest throughput for you?

I won't believe doing an extra round trip through a gen_* server
is faster than pushing a static chunk of binary data to a list of  
sockets.

Please prove me wrong!


> Also if you do a tcp send from different processes to the same socket
> you will get contention, and no guaranteed order, just an FYI.

No such use case.

> The whole reason for me using Erlang was that I could do one process
> per socket, as processes were so cheap.


I agree with you on processes being cheap and I use processes liberally.
I'm trying to optimize a specific use case, though, which is pushing
a chunk of binary data to 20k clients.

Also, I spent a LOT of time optimizing OpenPoker last fall. I used to  
have
an architecture where everything was a process. I had a gen_server for  
the
common game functionality, a gen_fsm for game logic, gen_servers for  
handling
pots, seats, etc.

Drastically cutting down on the number of processes significantly  
improved throughput.

---
Mac hacker with a performance bent
http://www.linkedin.com/in/joelreymont