Yes, you use SimPoint with the “fast” model (something simple, with no timing) so that the tool knows the activity. It then determines windows of importance so that you can run only those windows with the cycle accurate model.
From: Markus Bichl via gem5-users <gem5-users@gem5.org> Date: Wednesday, November 9, 2022 at 3:49 PM To: gem5-users@gem5.org <gem5-users@gem5.org> Cc: Markus Bichl <e1625...@student.tuwien.ac.at> Subject: [gem5-users] Re: SPEC CPU 2017 taking days to simulate Hi Jason, hi Jonathan, thanks a lot for pointing out SimPoint! I’m currently searching for details on how to use SimPoint with gem5. Is it right that I need to run the benchmarks at first using a timing CPU to take checkpoints at first? I guess ZjQcmQRYFpfptBannerStart This Message Is From an External Sender ZjQcmQRYFpfptBannerEnd Hi Jason, hi Jonathan, thanks a lot for pointing out SimPoint! I’m currently searching for details on how to use SimPoint with gem5. Is it right that I need to run the benchmarks at first using a timing CPU to take checkpoints at first? I guess (and hope) there is a way to generate the Simpoints in a faster way. Can someone give me a hint, or a maybe some kind of guidance on how to use Simpoint with gem5 v22? Thanks a lot! BR, Markus On 08.11.2022, at 20:10, Jonathan Kang <mos...@meta.com<mailto:mos...@meta.com>> wrote: Something as big as Spec typically wouldn’t be simulated as a whole but rather, using SimPoint on a cycle-accurate model. From: Jason Lowe-Power via gem5-users <gem5-users@gem5.org<mailto:gem5-users@gem5.org>> Date: Tuesday, November 8, 2022 at 11:05 AM To: The gem5 Users mailing list <gem5-users@gem5.org<mailto:gem5-users@gem5.org>> Cc: Markus Bichl <e1625...@student.tuwien.ac.at<mailto:e1625...@student.tuwien.ac.at>>, Jason Lowe-Power <ja...@lowepower.com<mailto:ja...@lowepower.com>> Subject: [gem5-users] Re: SPEC CPU 2017 taking days to simulate Hi Markus, I would expect gem5 to be at least 10,000-100,000x slower than your host. So, if it takes 100 seconds on the host, then I would expect between 1,000,000 and 10,000,000 seconds or more! That's 277-2770 hours or 10-100 days!! ZjQcmQRYFpfptBannerStart This Message Is From an External Sender ZjQcmQRYFpfptBannerEnd Hi Markus, I would expect gem5 to be at least 10,000-100,000x slower than your host. So, if it takes 100 seconds on the host, then I would expect between 1,000,000 and 10,000,000 seconds or more! That's 277-2770 hours or 10-100 days!! BTW, I actually think a 10-100,000x slowdown is on the low side if you're simulating multiple cores and/or using the out-of-order CPU model. In other words, running ref in gem5 to completion is not feasible :). There are techniques like SimPoint and other sampling methodologies that can help, but they come with tradeoffs, too. We're working to get some resources available to easily use sampling methodologies as well. Cheers, Jason On Tue, Nov 8, 2022 at 10:19 AM Markus Bichl via gem5-users <gem5-users@gem5.org<mailto:gem5-users@gem5.org>> wrote: Dear gem5 users, I’m currently running SPEC CPU 2017 benchmarks. I was succesful creating a SPEC CPU 2017 disk image from this tutorial: https://gem5.googlesource.com/public/gem5-resources/+/refs/heads/stable/src/spec-2017/<https://gem5.googlesource.com/public/gem5-resources/+/refs/heads/stable/src/spec-2017/> I used this image to start 3 benchmark runs on 3 different systems (server instance, desktop computer, and my laptop) 2 days ago. All of the systems have Intel Core/Xeon CPUs (but 5-7 years old), and 16 GBytes of DDR4 RAM. At this time, CPU load is 100% on one core for each system, at least 5 Gbytes of free memory. I started benchmarks using these commands: ./build/X86/gem5.fast ./configs/example/gem5_library/x86-spec-cpu2017-benchmarks.py --image /home//Projects/spec-2017/spec-2017-image/spec-2017 --partition 1 --benchmark 625.x264_s --size ref ./build/X86/gem5.fast ./configs/example/gem5_library/x86-spec-cpu2017-benchmarks.py --image /home//Projects/spec-2017/spec-2017-image/spec-2017 --partition 1 --benchmark 600.perlbench_s --size ref ./build/X86/gem5.fast ./configs/example/gem5_library/x86-spec-cpu2017-benchmarks.py --image /home//Projects/spec-2017/spec-2017-image/spec-2017 --partition 1 --benchmark 502.gcc_r --size test I wanted to have the results as fast as possible, so I used gem5.fast, and also tried the gcc benchmark with the test workload. As I do not see any progess on the benchmarks yet, I feel there is something wrong. Is this a normal simulation time for SPEC CPU 2017 benchmarks? Is it possible to track the progress of the simulations, even with "m5.disableAllListeners()” enabled in the system configuration? A run of 600.perlbench_s using the ref workload directly on the host system (server instance) took 181 seconds. Thanks a lot for your help! BR, Markus Bichl Student, Technische Universität Wien _______________________________________________ gem5-users mailing list -- gem5-users@gem5.org<mailto:gem5-users@gem5.org> To unsubscribe send an email to gem5-users-le...@gem5.org<mailto:gem5-users-le...@gem5.org>
_______________________________________________ gem5-users mailing list -- gem5-users@gem5.org To unsubscribe send an email to gem5-users-le...@gem5.org
