Hi Rizwana, The KVM CPU is really for fast forwarding. When you use it, you will get no stats, and no parts of the memory system are exercised.
You could envision building up a sampling-based methodology. In fact, there has been some work done already, as presented at the ISCA gem5 workshop. Andreas From: gem5-users <[email protected]<mailto:[email protected]>> on behalf of Rizwana Begum <[email protected]<mailto:[email protected]>> Reply-To: gem5 users mailing list <[email protected]<mailto:[email protected]>> Date: Wednesday, 5 August 2015 20:35 To: gem5 users mailing list <[email protected]<mailto:[email protected]>> Subject: Re: [gem5-users] Reducing time taken by dual simulations Hello Andreas, We are considering to use ArmKvmCPU. I have a couple of questions regarding its simulation methodology. From my understanding so far, KVM is a good way to speedup simulations by switching between virtual (faster) and simulated CPU. So, I am wondering if we can rely on all Gem5 stats that will be reported after the simulation? Will running an arm_detailed CPU and ArmKvmCPU result in similar stats? Also, does KVM CPU use dram controller from Gem5 to simulate memory or it uses host physical memory? Thanks, -Rizwana On Mon, Aug 3, 2015 at 3:31 PM, Davesh Shingari <[email protected]<mailto:[email protected]>> wrote: Thanks a lot for clarification. I didn't know about that. [https://mailfoogae.appspot.com/t?sender=ac2hpbmdhcmlkYXZlc2hAZ21haWwuY29t&type=zerocontent&guid=6881708c-88a7-4e14-877f-a79820b74647]ᐧ On Mon, Aug 3, 2015 at 9:47 AM, Guru Prasad <[email protected]<mailto:[email protected]>> wrote: Gem5 has a command line argument --dual. This starts up 2 separate simulations (inside 1 process) where one is called the driver and the other is the 'real' test system. Since networking requires at least two systems, you usually use --dual when you want to test the network. Because gem5 is not yet multi-threaded (I think), both simulations end up using only 1 core. Hence, ideally this should result in a 2x slowdown. However, I'm seeing a lot more. That's what my question was about. My single system simulations also take about 12-15 hours to complete. With detailed they take a lot longer (haven't fully completed a sim yet). Regards Guru On Mon, Aug 3, 2015 at 12:42 PM, Davesh Shingari <[email protected]<mailto:[email protected]>> wrote: Just to be sure,what do you mean by dual? Do you mean 2 simulation or multiple core configuration? [https://mailfoogae.appspot.com/t?sender=ac2hpbmdhcmlkYXZlc2hAZ21haWwuY29t&type=zerocontent&guid=d2c7bc4c-1b48-416c-8b33-eac5f0965824]ᐧ On Mon, Aug 3, 2015 at 9:39 AM, Guru Prasad <[email protected]<mailto:[email protected]>> wrote: Hi Davesh, just to clarify, is this with --dual? Regards Guru On Mon, Aug 3, 2015 at 12:38 PM, Davesh Shingari <[email protected]<mailto:[email protected]>> wrote: Hi Guru I am using Android on Gem5 for my work. When I use the simulation to take checkpoint, then it takes around 12 - 15 hours to get checkpoint and finish simulation. The simulation time is 32 seconds after the checkpoint. But when I use that checkpoint along with detailed mode configuration and full memory system, then it takes almost a week to complete. [https://mailfoogae.appspot.com/t?sender=ac2hpbmdhcmlkYXZlc2hAZ21haWwuY29t&type=zerocontent&guid=faa2c819-ef06-4300-b7a3-22a83834a58a]ᐧ On Mon, Aug 3, 2015 at 9:23 AM, Guru Prasad <[email protected]<mailto:[email protected]>> wrote: Ah I see. It sounds like the sort of like the hack we made for hdlcd where we read from physical memory directly. Thanks once again. I will try the options you've described. Regards Guru On Mon, Aug 3, 2015 at 12:19 PM, Andreas Hansson <[email protected]<mailto:[email protected]>> wrote: > Hi Guru, > > 3 days for 20 seconds sounds way slower than it should be. > > Fast-mem is not replacing the DRAM controller, it bypasses the entire > memory system. In atomic we do not do any sensible timing simulation of > the controller in any case. > > The KVM CPU is only built if your system support it. Hence, you need to > build gem5 for ARM on an ARM system. A good reason to get an ARM server :-) > > Andreas > > On 03/08/2015 16:46, "gem5-users on behalf of Guru Prasad" > <[email protected]<mailto:[email protected]> on behalf of > [email protected]<mailto:[email protected]>> wrote: > >>Hi Andreas, >> >>Thanks for the suggestions. It took about 3 days to simulate 20seconds >>under atomic. >>I don't know how to tell if interrupts are being lost. The kernel is >>not complaining about anything on kmsg. >>I will try atomic with --fast-mem. Currently, I've been using >>lpddr3_1600_x32 >> >>I don't think I have KvmArmCPU available. I will try the same with the >>latest. >>Is using KvmArmCPU as simple as just saying --cpu-type=...? >>I ask this because --list-cpu-types for my current version only displays >> timing => TimingSimpleCPU >> detailed => DerivO3CPU >> atomic => AtomicSimpleCPU >> minor => MinorCPU >> >> >>Regards >>Guru >> >>On Mon, Aug 3, 2015 at 11:33 AM, Andreas Hansson >><[email protected]<mailto:[email protected]>> wrote: >>> Hi Guru, >>> >>> To speed things up, here are some options: >>> >>> 1. Use atomic to get the point of interest, then take a checkpoint. Is >>> this what you are doing? >>> >>> 2. For the above, run with --fast-mem since you’re not warming any >>>caches >>> anyways. >>> >>> 3. If possible, use the KvmArmCPU instead of (1). This should give you a >>> significant boost. >>> >>> If the amount of work is constant, than running with dual should not >>>slow >>> things down. If you’re seeing a slow down, then perhaps something odd is >>> going on in the simulation (interrupts getting lost etc). Have you tried >>> with the latest trunk? >>> >>> Andreas >>> >>> >>> On 03/08/2015 16:16, "gem5-users on behalf of Guru Prasad" >>> <[email protected]<mailto:[email protected]> on behalf >>> of [email protected]<mailto:[email protected]>> wrote: >>> >>>>Hi, >>>> >>>>Recently we've been trying to run some experiments with the network. >>>>To do so, we've been using dual simulations. So far, we've been able >>>>to get away with running two simulations booting barebones linux with >>>>just a shell. However, our real workload is BBench and so we need to >>>>boot a full Android image. >>>> >>>>A single simulation of full Android takes us about ~12hours to boot. >>>>However under dual, its been about 3 days and has so far simulated >>>>only ~20seconds. Is there any way to speed up the simulation? Is there >>>>any way we can stitch two single simulations together (initial loss of >>>>time on drivesys is acceptable). >>>> >>>>Currently we're on revision e179497451. >>>> >>>>Regards >>>>Guru >>>>_______________________________________________ >>>>gem5-users mailing list >>>>[email protected]<mailto:[email protected]> >>>>http://m5sim.org/cgi-bin/mailman/listinfo/gem5-users >>> >>> >>> -- IMPORTANT NOTICE: The contents of this email and any attachments are >>>confidential and may also be privileged. If you are not the intended >>>recipient, please notify the sender immediately and do not disclose the >>>contents to any other person, use it for any purpose, or store or copy >>>the information in any medium. Thank you. >>> >>> ARM Limited, Registered office 110 Fulbourn Road, Cambridge CB1 9NJ, >>>Registered in England & Wales, Company No: 2557590 >>> ARM Holdings plc, Registered office 110 Fulbourn Road, Cambridge CB1 >>>9NJ, Registered in England & Wales, Company No: 2548782 >>> _______________________________________________ >>> gem5-users mailing list >>> [email protected]<mailto:[email protected]> >>> http://m5sim.org/cgi-bin/mailman/listinfo/gem5-users >>_______________________________________________ >>gem5-users mailing list >>[email protected]<mailto:[email protected]> >>http://m5sim.org/cgi-bin/mailman/listinfo/gem5-users > > > -- IMPORTANT NOTICE: The contents of this email and any attachments are > confidential and may also be privileged. If you are not the intended > recipient, please notify the sender immediately and do not disclose the > contents to any other person, use it for any purpose, or store or copy the > information in any medium. Thank you. > > ARM Limited, Registered office 110 Fulbourn Road, Cambridge CB1 9NJ, > Registered in England & Wales, Company No: 2557590 > ARM Holdings plc, Registered office 110 Fulbourn Road, Cambridge CB1 9NJ, > Registered in England & Wales, Company No: 2548782 > _______________________________________________ > gem5-users mailing list > [email protected]<mailto:[email protected]> > http://m5sim.org/cgi-bin/mailman/listinfo/gem5-users _______________________________________________ gem5-users mailing list [email protected]<mailto:[email protected]> http://m5sim.org/cgi-bin/mailman/listinfo/gem5-users -- Have a great day! Thanks and Warm Regards Davesh Shingari Master's in Computer Engineering [EE] Arizona State University [email protected]<mailto:[email protected]> _______________________________________________ gem5-users mailing list [email protected]<mailto:[email protected]> http://m5sim.org/cgi-bin/mailman/listinfo/gem5-users _______________________________________________ gem5-users mailing list [email protected]<mailto:[email protected]> http://m5sim.org/cgi-bin/mailman/listinfo/gem5-users -- Have a great day! Thanks and Warm Regards Davesh Shingari Master's in Computer Engineering [EE] Arizona State University [email protected]<mailto:[email protected]> _______________________________________________ gem5-users mailing list [email protected]<mailto:[email protected]> http://m5sim.org/cgi-bin/mailman/listinfo/gem5-users _______________________________________________ gem5-users mailing list [email protected]<mailto:[email protected]> http://m5sim.org/cgi-bin/mailman/listinfo/gem5-users -- Have a great day! Thanks and Warm Regards Davesh Shingari Master's in Computer Engineering [EE] Arizona State University [email protected]<mailto:[email protected]> _______________________________________________ gem5-users mailing list [email protected]<mailto:[email protected]> http://m5sim.org/cgi-bin/mailman/listinfo/gem5-users -- IMPORTANT NOTICE: The contents of this email and any attachments are confidential and may also be privileged. If you are not the intended recipient, please notify the sender immediately and do not disclose the contents to any other person, use it for any purpose, or store or copy the information in any medium. Thank you. ARM Limited, Registered office 110 Fulbourn Road, Cambridge CB1 9NJ, Registered in England & Wales, Company No: 2557590 ARM Holdings plc, Registered office 110 Fulbourn Road, Cambridge CB1 9NJ, Registered in England & Wales, Company No: 2548782
_______________________________________________ gem5-users mailing list [email protected] http://m5sim.org/cgi-bin/mailman/listinfo/gem5-users
