[RC5] Windows Stream client doesnt like the normal client

Lars Sahlmann lanisa at gmx.de
Fri May 14 05:34:05 EDT 2010


thanks for the information. Speed decrease roughly 5% on my i7 rig running together with ogr.I always thought the second card dropped my top speed. 

Anyway HT gives a nice additional boost on OGR computation. iirc 25% on that rig. 




-------- Original-Nachricht --------
> Datum: Thu, 13 May 2010 15:58:40 -0700
> Von: Gerald Richter <glassman1 at neonshadow.net>
> An: "D.net Discussion" <rc5 at lists.distributed.net>
> Betreff: Re: [RC5] Windows Stream client doesnt like the normal client

> The stream and cuda clients both require 1 cpu thread free to feed them 
> in order to maintain top speed. I run the CPU client configured with 
> real CPU cores -1 parallel to the stream and cuda clients in my 
> environments.
> 
> Jonathan Dixon wrote:
> > I ran into the same problem running the CUDA client + normal client.  I 
> > have a GeForce GTX260 and Core i7.  I noticed that running the normal 
> > client would slow down the CUDA client enough that it wasn't worth 
> > running the normal client at all, but I hated to let a Core i7 sit 
> > virtually idle.  I had tried to increase the priority of the GPU client,
> > but it didn't seem to help any.  The solution I came up with was to 
> > manually set the core count on the normal client to one less than the 
> > maximum number of threads on the computer.  This prevented processes 
> > from moving from core to core, which slows them down a little, and also 
> > should prevent competition between the clients because there is always a
> > free core to keep feeding the GPU. 
> > 
> > I also agree with Waldo on the hyper-threading.  HT is great for single 
> > core CPUs and helps them multi-task better, but it has a negligible 
> > effect when the computer is already good at multi-tasking due to the 
> > number of cores.  Given that a little efficiency is lost when the 
> > computer moves processes to and from a specific core, it is counter 
> > productive to run two jobs on the same core that accomplish the same 
> > task as does HT.  HT is designed to reduce the effect, but if the 
> > computer is built to run RC5, it would be better just to not have the 
> > effect at all.  If RC5 is set to auto-detect thread count and HT is 
> > enabled, it will automatically assign two processes to every core.  Thus
> > HT would seem to degrade performance by inducing exactly what it is 
> > supposed to solve.  I don't think it is effective to leave HT running 
> > and manually set the cores to half because the the core time has already
> > been split and you are deciding to only use half of it.  Yes, I know. an
> > 8 threaded computer just isn't as cool as a 16 threaded computer.
> > 


-- 
GRATIS für alle GMX-Mitglieder: Die maxdome Movie-FLAT!
Jetzt freischalten unter http://portal.gmx.net/de/go/maxdome01


More information about the rc5 mailing list