Re: [Gems-users] GEMS Simulator Simulation time

[Dan Gibson <degibson@xxxxxxxx>]

Hrishikesh,

Speeding up memory timing models is difficult. Basically, Simics is a 
cycle-by-cycle simulator, bu timing models cause the Simics processors 
to stall. When a Simics processor stalls, it seems to still be doing 
some work, simply because logical time advances. When memory timing 
models stall processors, most processors are stalled for most of the time.
Using Simics (again, this is actually a Simics issue), there is no clear 
way to get around this if you care about interprocessor timing 
dependencies. However, if you *only* care about cache miss rates and NOT 
the effect of one processor progressing slower than another due to cache 
effects, there may be a solution that will speed your simulations: Drop 
the cache access latencies to near-zero, to minimize the time that 
Simics is stalled. This will definately cause execution behaviors *not* 
consistent with real memory timings, but it will also reduce the amount 
of time the simulation is stalled on memory accesses.
You can change access latencies in ruby/config/rubyconfig.defaults. Then 
recompile Ruby.
As far as benchmarks, I recommend you write some code of your own and 
run it using both Simics only and Simics+Ruby configurations, since I 
don't know what benchmarks you are using.
Good luck.

Regards,
Dan

Hrishikesh Das wrote:

> Dan,
>
> Thanks for the prompt reply,
>
> Is there a way I can speed it up? Right now I am not 
> loading/simulating the Opal module. Will loading this result in more 
> slowdown(as I expect) ?
> Also as an alternative whats the simplest benchmark that I can test ? 
> I mean a very simple benchmark that executes very fast.
> Thanks,
> Hrishikesh.
>
>
> On 11/2/05, *Dan Gibson* <degibson@xxxxxxxx 
> <mailto:degibson@xxxxxxxx>> wrote:
>     Hrishikesh,
>
>     The performance impact of using a memory timing model (like Ruby) is
>     very significant. The slowdowns you are reporting are not unusual--in
>     some cases running Simics with Ruby can slow simulation very
>     significantly.
>
>     The reason for the slowdown is quite complex, and it not accounted for
>     entirely by execution within the timing model. Specifically, Simics
>     itself runs (much) slower when memory timing models are used.
>
>     Regards,
>     Dan Gibson
>
>     Hrishikesh Das wrote:
>
>     > Dear Sirs,
>     >
>     > I am a student using GEMS in a Parallel computing architecture
>     course.
>     > I am using SIMICS and a SIMICS suse 7.3 image running on a X86 dual
>     > processor server.
>     >
>     > The problem I am facing is that when using Ruby for simulation, the
>     > simulation is extremely slow. Ex- a simple simulation of Unixbech,
>     > that takes SIMICS about 1 -2 minutes to emulate, takes SIMICS with
>     > Ruby about 10+ hours. Is this kind of slowdown natural ? Is there
>     > anything I can do to speed things up ? I need to get these results
>     > done for my course work,
>     >
>     > Any help will be greatly appreciated.
>     > Thanks,
>     > Hrishikesh Das.
>     >
>     > Graduate Student,
>     > Mississippi State University.
>     >
>     >------------------------------------------------------------------------
>     >
>     >_______________________________________________
>     >Gems-users mailing list
>     >Gems-users@xxxxxxxxxxx <mailto:Gems-users@xxxxxxxxxxx>
>     >https://lists.cs.wisc.edu/mailman/listinfo/gems-users
>     <https://lists.cs.wisc.edu/mailman/listinfo/gems-users>
>     >
>     >
>     _______________________________________________
>     Gems-users mailing list
>     Gems-users@xxxxxxxxxxx <mailto:Gems-users@xxxxxxxxxxx>
>     https://lists.cs.wisc.edu/mailman/listinfo/gems-users
>
>
> ------------------------------------------------------------------------
>
> _______________________________________________
> Gems-users mailing list
> Gems-users@xxxxxxxxxxx
> https://lists.cs.wisc.edu/mailman/listinfo/gems-users
>