Use pinpoints tool for SPEC benchmarks simpoints generation. Dont use Gem5 for BBV file generation.
-Hassan On Wed, Jun 28, 2017 at 11:59 PM, Oscar Rosell <[email protected]> wrote: > Hi, > > I was not really suggesting using SimPoints (even that is a good option > depending on your needs). As you say, to generate the SimPoints you first > have to run the full thing (although you can do it with less detailed CPU > which greatly speeds up things). I was just stating the typical throughput > (simulated instructions / real time) we're getting from Gem5 with O3 CPU so > that James can check if his performance is "normal". > > I have never run perlbench but I have run many workloads taking several days > to finish so 3 days doesn't look too bad to me. > > Regards, > > Oscar > > > On 28/06/17 15:55, Asif Ali Khan wrote: > > Hi Oscar, > > Yes, I agree with Simpoints its possible to simulate SPEC benchmarks but the > problems is, how do you generate BBV file for the simpoint tool. I tried to > generate this BBV file using Gem5 (for perlbench) and took like 3 days. Did > I do something wrong? > > Asif > > > On Wednesday, June 28, 2017, 3:01:43 PM GMT+2, Oscar Rosell > <[email protected]> wrote: > > > Hi, > > Just so that you have a reference, we usually simulate SimPoints of 100M > instructions (with 10M extra warmup instructions) and on a single core > O3+L1+L2 model they take between 30 minutes and 75 minutes. > > Regards, > > Oscar > > > On 28/06/17 10:47, Stine, James wrote: >> Hi All, >> >> I am sorry to bother everyone. I am trying to gauge performance and would >> love some feedback on run-time performance. My main impetus for this Email >> is due to limited information I could find and just wanted to get some >> feedback on if there was some issues related to this topic. I apologize in >> advance if I missed something specific about this question. >> >> I did some tests on the queens benchmark as well as some others and my run >> times seem to take a long time. 16X grids within queens.c (e.g., queens 16) >> seem to run about 17 hours using AtomicMemory access with caching. The >> ASPLOS-13 tutorial seems to have very small numCycles, so not sure that is >> accurate for “-o 16” on queens.c. Eventually, I would love SPEC, but I am >> quite worried if queens.c takes forever, how can I even manage to get SPEC >> through. I also tried some other benchmarks like Matrix Multiplications, >> but some of them take just as long. However, queens does take a while to >> run, which I know is typical due to its intense computation mix. My x86 >> cycle counts (statically compiled with -O3 and loop unrolling) were: >> 60,055,907,458 on multi-core Intel extreme processors - again, I might have >> not run something correctly. >> >> If anyone can possibly share their tips/tricks - especially for eventual >> running of SPEC, it would be great. Does anyone do anything to maximize >> performance? Even the smallest of tips would be helpful. Perhaps, I am >> running gem5 with the wrong settings. Or, perhaps, the settings are correct >> and this is a normal set of run times. Anyways, I appreciate any help and >> also appreciate the wonder of gem5. Take care. >> >> All my best, >> >> James >> _______________________________________________ >> gem5-users mailing list >> [email protected] >> http://m5sim.org/cgi-bin/mailman/listinfo/gem5-users > > _______________________________________________ > gem5-users mailing list > [email protected] > http://m5sim.org/cgi-bin/mailman/listinfo/gem5-users > > > _______________________________________________ > gem5-users mailing list > [email protected] > http://m5sim.org/cgi-bin/mailman/listinfo/gem5-users > > > > _______________________________________________ > gem5-users mailing list > [email protected] > http://m5sim.org/cgi-bin/mailman/listinfo/gem5-users _______________________________________________ gem5-users mailing list [email protected] http://m5sim.org/cgi-bin/mailman/listinfo/gem5-users
