It's been a while since I've looked at the code, but I believe Nate has most
of it correct. I think commit.
COM:count should include most all instructions. committedInsts most likely
excludes noops. I don't remember the exact reason for the disjoint between
the sum of all counts and sim_insts, but I believe there may be some
counting issues with PAL code instructions. I think there are some issues
with how legacy code handled cetain types of instructions, leading to these
funny counts. Definitely things that should be commented better in the
future.

Kevin

On Fri, Aug 6, 2010 at 2:20 PM, nathan binkert <[email protected]> wrote:

> I think it mostly has to do with two things.  First, I don't believe
> that we count noops in the committed instructions.  Second, (and I'm
> not sure of this), but sim_insts may count misspeculated instructions.
>
> In terms of what you need to use, you really need to ask the McPAT authors.
>
> I'll be honest, our stats could use a bit of an overhaul.  The naming
> and explanation of what they all are is confusing (not unlike real CPU
> performance counters :)  If anyone wants to work on that (and you
> could just do it on a particular object), we'd welcome the effort.
>
>  Nate
>
> On Thu, Aug 5, 2010 at 8:30 AM, Iordan Alexandru <[email protected]> wrote:
> >
> > Hello
> >
> > I am a new user of M5 so I am sorry if the following will sound trivial.
> >
> > I want to do some full-system performance/power estimations so I am
> running the PARSEC benchmarks in M5 (for performance measurements) and then
> pass the results to McPAT (for power and area estimations). My problem is
> that I don't know which ones of M5's stats to transfer to McPAT, in the sens
> that there are many who appear to point to the same thing. Some examples:
> >
> > system.switch_cpus1.commit.COM:count           146267
>     # Number of instructions committed
> >
> > system.switch_cpus1.commit.COM:committed_per_cycle::total       238806
>                     # Number of insts commited each cycle
> >
> > system.switch_cpus1.committedInsts             143430
>   # Number of Instructions Simulated
> >
> > sim_insts                                   118094785
>   # Number of instructions simulated
> >
> > (I should mention that if you add the system.switch_cpusX.committedInsts
> for all CPUs you don't get the sim_insts)
> >
> > I also find it a bit strange that if you add all the different types of
> committed instructions that are reported, you don't end up with the
> committed.COM:count value.
> >
> > Does anyone has a better idea about (or some experience on working with)
> this counters?
> >
> > 10x in advance
> >
> >
> >
> > _______________________________________________
> > m5-users mailing list
> > [email protected]
> > http://m5sim.org/cgi-bin/mailman/listinfo/m5-users
> >
> >
> _______________________________________________
> m5-users mailing list
> [email protected]
> http://m5sim.org/cgi-bin/mailman/listinfo/m5-users
>
_______________________________________________
m5-users mailing list
[email protected]
http://m5sim.org/cgi-bin/mailman/listinfo/m5-users

Reply via email to