[RC5] Partly finished blocks.

Jeff Lawson jlawson at bovine.net
Wed Dec 4 15:39:37 EST 2002


On Wed, 4 Dec 2002, Dennis Lubert wrote:

> I just downloaded the new RC5-72 client for windows (.477), and got some
> problems with partly finished blocks. I stopped and started the client
> several times, but sometimes I get this error message : Packet was from a
> different core/client cpu/os/build. And the block is discarded. I further
> noticed in the logs that the client uses sometimes core #0 and sometimes #4
> on my PIII. Is this the reason ? And why is the client sometimes using
> another core ?
>

One thing that may be partly related is the fact that hardcoded core
selections based on the cpuid are currently disabled, and the client
always performs a microbench at startup to determine which core is the
best performing on your hardware.  This is because the performance
characteristics of all of the rc5-72 cores are still very much in flux
while they are still being optimized and new cores created, so it is very
hard to keep a hard-coded list of "which core is best for cpuid X" up to
date.  (This will change in future versions when the cores are not
changing as much anymore.)

If the performance difference between two cores on your particular
hardware is very miniscule, it may indeed turn out that the client may
probablistically select a different one on each run.  The accuracy of the
microbench may also be slightly impacted by other activity on your machine
at the time the core selection is being performed.

The client will automatically restart a checkpointed workunit from the
beginning if it was processed using a core different from the one that it
has selected as being the optimal one for your processor.  This is
definitely not ideal (and not always necessary), but it is difficult to
guarantee identical block resumption results when there are cores that use
different key incrementing patterns.

As other users have pointed out, the new client is also incapable of
reading workunits or checkpoint files created by prior v2.8 versions of
the client.  However in your case it sounds like your checkpoint file is
indeed being created by the new version and you're just having trouble
resuming from it.  If this is indeed your circumstance and the
core-selection issue is the cause of the block restarts, you may wish to
consider setting the client to always use a fixed core number, instead of
automatically selecting one with a microbench.  This option is under the
"Performance related options" page in the client configuration.


-- 
Jeff Lawson   http://www.bovine.net/  http://www.distributed.net/~bovine/
Jeffrey_Lawson at alumni.hmc.edu  jlawson at bovine.net  bovine at distributed.net
Programmer, Developer, Mascot, Founder of the largest computer on earth!
Don't waste those cycles!  Put them to use!  http://www.distributed.net/


--
To unsubscribe, send 'unsubscribe rc5' to majordomo at lists.distributed.net
rc5-digest subscribers replace rc5 with rc5-digest



More information about the rc5 mailing list