Hi, I am running parsec on X86, using gem5-stable-07352f119e48. I am using the precompiled Parsec binaries and kernel images. I use 2 cores and 2 threads.I restore from a checkpoint (which was taken at the beginning of the ROI). I dumpreset statistics every total 10 million instructions (CPU0 + CPU1). However, for some benchmarks such as facesim, streamcluster, freqmine, the IPS values are too imbalanced. For example for streamcluster, while CPU0 has 1400 million instructions per second, CPU1 has 0.4 million instructions per second. I calculate IPS myself by dividing switch_cpus.committedInsts by the sim_seconds for that stat. I see the similar imbalance for 6 core 6 thread case even for blackscholes which should be fairly clean. I also see that the IPS and IPC traces over time is weirdly flat for some other benchmarks such as dedup and x264 which were reported to have lots of variations. I tried other ways to dump statistics such as using the built in function statDump, but I observed flat characteristics for that case, too. Could anyone help me with this issue? Could it be related to restoring from the checkpoint (which i do not get any errors about) or dumping statistics or something else? I can provide some IPC, IPS trace plots for benchmarks if needed. Best, Fulya Kaplan
_______________________________________________ gem5-users mailing list [email protected] http://m5sim.org/cgi-bin/mailman/listinfo/gem5-users
