FW: Re: [Spread-users] Spread Performance

[John Schultz]

Ciprian Tutu wrote:
> Hi,
> 
> WI> I configured 3 machines, over 100Mbit Ethnet, each PIIIs running Windows XP
> WI> (yes, I know evil Microsoft). 2 machines I configured with Spread Daemons
> WI> with apps I wrote implimenting the Spread API to receive messages. 1 machine
> WI> served as a message sender and connected to one of the other machines Spread
> WI> Daemons.
> 
> Honestly I would configure one daemon on each of your 3 machines and have
> each client application (including the sender) connect to their local
> daemon. This way you really let spread handle _all_ the traffic of the
> experiment.
> 
> Add your 3rd daemon to the config file on each machine and give it a
> try.
> 

I agree with this recommendation.  You really want Spread running on 
every machine with processes that could call SP_connect.  Usually your 
SP_connect call should be a hardcode to <port_num>@localhost.

Another thing to check is whether or not your configuration is losing 
messages at all.  If the network is losing tokens intermittently but 
often this would slow you down.  You can detect this by examining the 
state of all the daemons using spmonitor and looking at the retrans 
fields.  If these numbers are in the hundreds or thousands, then you 
have a lossy network or NICs and need to fix that.

Regardless, in recent experiments that we have run on Spread using 
Windows in small configurations we couldn't do much better than 40Mb 
(5MB) / sec.  We believe that Spread's event system on Windows' doesn't 
perform as well as it does on *nix.  We were trying to fix this before 
other work intervened.

-- 
John Lane Schultz
Spread Concepts LLC
Phn: 443 838 2200