Re: [Gems-users] L2 Miss Profiling

[Dan Gibson <degibson@xxxxxxxx>]

Ahh, ocean. If only I had a dollar for every time I've seen this same issue with ocean. Here is what is going on:

258x258 elements * 4 bytes/integer * 2 copies of the ocean (because of shadowing) =
520 KB working set size.

Aggregate L1 cache size = 64KB * 16 Processors = 1MB.
Aggregate L2 cache size = 1MB.

In other words, after the first full timestep, the entire ocean is resident in the L1 caches. Any misses from the L1s at that point are sharing misses, which 'miss' in the L2 because other L1s must be invalidated.

Also, for this particular case, the L2 is no bigger than the aggregation of the L1s. I would hope that the L2 isn't inclusive, but I honestly don't know if it is or not for MESI_CMP_filter_directory. If it IS inclusive, obviously an L2 hit would only happen rarely, on a transient condition.

Regards,
Dan

On Thu, Apr 8, 2010 at 11:44 AM, Nilay Vaish <nilay@xxxxxxxxxxx> wrote:

On Thu, 8 Apr 2010, Dan Gibson wrote:

It is not out of the question that there might be profiling problems with
MESI_CMP_filter_directory. However, lets cover the basics first:
- What are your cache sizes?

L2  cache_size_Kbytes: 1024
L1I cache_size_Kbytes: 64
L1D cache_size_Kbytes: 64

- What is the workload?

I am simulating the Ocean application with 4 threads. It makes use of a 258 X 258 grid of integers.

- For how long are you running the simulation?

The simulation runs for about 18 minutes.

Regards,
Dan

On Thu, Apr 8, 2010 at 10:52 AM, Nilay Vaish <nilay@xxxxxxxxxxx> wrote:

Hi

I am using GEMS for studying effect of L2 cache parameters on performance
of CMPs. Right now I am using the protocol MESI_CMP_filter_directory. The L2
misses obtained as output is suspiciously close to the sum of L1D and L1I
cache misses. I am quoting the misses obtained for a 16 processor machine.

L1D_cache cache stats:
 L1D_cache_total_misses: 1334256
 L1D_cache_total_demand_misses: 1334256

L1I_cache cache stats:
 L1I_cache_total_misses: 2010
 L1I_cache_total_demand_misses: 2010

L2_cache cache stats:
 L2_cache_total_misses: 1336262
 L2_cache_total_demand_misses: 1336262

Has anyone used GEMS for profiling L2 misses? What might be going on wrong
here?

--
Nilay

_______________________________________________
Gems-users mailing list
Gems-users@xxxxxxxxxxx
https://lists.cs.wisc.edu/mailman/listinfo/gems-users
Use Google to search the GEMS Users mailing list by adding "site:
https://lists.cs.wisc.edu/archive/gems-users/" to your search.

--
http://www.cs.wisc.edu/~gibson [esc]:wq!

Thanks

Nilay
_______________________________________________
Gems-users mailing list
Gems-users@xxxxxxxxxxx
https://lists.cs.wisc.edu/mailman/listinfo/gems-users
Use Google to search the GEMS Users mailing list by adding "site:https://lists.cs.wisc.edu/archive/gems-users/" to your search.

-- 
http://www.cs.wisc.edu/~gibson [esc]:wq!