2

I'm looking to use gem5 with some of my work, and have a very general question regarding its features.

My question is, using gem5, can I get statistics on behavior / system resource usage of individual threads, either in SE or FS mode. For example, if I have 2 threads running in my application, can I see that Thread 1 uses X cycles on CPU 1, Y cycles on CPU 2 etc, and Thread 2 uses Z cycles on CPU 3 etc. If this is possible, does it also extend to cache usage, memory usage etc?

I saw the following questions: Accessing logical (software) thread ID in gem5, but I think just using hardware thread statistics would work for our use case.

I see that the ThreadState class in src/cpu/thread_state.hh has the following feilds:

/** Number of instructions committed. */
Counter numInst;
/** Stat for number instructions committed. */
Stats::Scalar numInsts;
/** Number of ops (including micro ops) committed. */
Counter numOp;
/** Stat for number ops (including micro ops) committed. */
Stats::Scalar numOps;
/** Stat for number of memory references. */
Stats::Scalar numMemRefs;

This looks like something I could use. Would anyone know if these stats can be easily output per thread in the stats file?

Thanks!

Anthony Canino
  • 25
  • 2

1 Answers1

1

I haven't done this myself, but I will make the following educated guesses:

  • se.py: SE has a limitation that it can only run one thread per core, so I suspect that this will show directly on the per core stats

  • fs.py: the definition of thread is OS dependent, so you would need some kind of guest support.

    In the context of ARM Linux kernel tracing for example, CONFIG_PID_IN_CONTEXTIDR shows on logs: How to access logical software user thread ID of a Linux process in gem5? but not sure how easy it will be to get the stats out

So likely se.py will be the easier approach.

--enable-context-switch-stats-dump

I just came across this option while reading some source:

        parser.add_option("--enable-context-switch-stats-dump", \
                action="store_true", help="Enable stats dump at context "\
                "switches and dump tasks file (required for Streamline)")

It appears to hook then the PC hits the __switch_to kernel symbol address:

void
FsLinux::startup()
{
    FsWorkload::startup();

    if (enableContextSwitchStatsDump) {
        if (getArch() == Loader::Arm64)
            dumpStats = addKernelFuncEvent<DumpStats64>("__switch_to");
        else
            dumpStats = addKernelFuncEvent<DumpStats>("__switch_to");

so this should dump per thread for each context switch!

I haven't tested it however, and I wonder how easy it will be to correlate which thread is what.

Ciro Santilli
  • 3,693
  • 1
  • 18
  • 44