On Thu, Jan 07, 2010 at 05:29:49PM -0000, darren wrote: > I have just started crunching some tables out on a box with a nvidia GTX > 275 with the below parameters: > > > > ./a51table --condition rounds:rounds=32 --roundfunc > xor:condition=distinguished_point::bits=15:generator=lfsr2::tablesize=32 > ::advance=304256 --implementation sharedmem --algorithm A51 --device > cuda --operations 512 --work random:prefix=11,0 --consume > file:prefix=data:append --logger normal generate --chains 270000000 > --chainlength 3000000 --intermediate filter > > > > > > I am getting rate 170.00 chains.sec (interval: 00:01:00) >
after launch the number of chains/second goes a bit like a sinus wave, that is because all chains start at length zero and most of them need around 1M steps before ending all together at that point. because the length of the chains is a gaussian distribution, the sinus wave gets shallower and approaches a straight horizontal line lying on the x axis. in this representation the x axis is time and y is chains/second and the y axis zero line is the average chains/second. 170 is either because of that or because your card is clocked above 650 mhz. given the fact that you do 512 operations per kernel call i doubt that 170 is your average speed. how long did you run the program before getting the number? > > > Can you give me some tips on trying to squeeze out some more chains per > sec I think a GTX 275 shouls be able to crunch around 250 chains/sec no, at the upper limit of 700mhz it cannot produce more than 180 chains/sec. At least not with the old 4bit table lookup code, the new bitslice code is faster, but not yet stable. > > > > Is anyone else using a GTX 275 and if so can you let me know what > parameters you are using and what speeds you are getting? > you can increase operations up to around 8000, but your UI would become very sluggish then. you can overclock your GPU, but that can be tricky, as the GPU may start to calculate false values. To check whether this is the case, use --work increment ... --consume print generate --chains 16384, pipe stdout to a file and repeat and compare the output after changing the clock frequency. with --increment the start values are always the same and so should the end values be. > > > My other question is that if I stop the table generation and changes > some parameters and restart will it continue generating the tables from > where if stopped and just complete at a faster rate ? > --condition, the prefix parameter of --work, --roundfunc and the obvious --advance --algorithm must be the same for each invocation. i can only imaging changing the --operations parameter to adapt the speed /UI responsiveness or the --device parameter to continue on another GPU. and you would have to stop the program with $ a51table signal --shutdown and wait till it exits > > > And if it continues ok will the tables still be usable and valid when > completed ? > probably > > > Great to see alot more activity on the list since C26 also J > we reached 400 subscribers yesterday. _______________________________________________ A51 mailing list [email protected] http://lists.lists.reflextor.com/cgi-bin/mailman/listinfo/a51
