On Wed, 15 May 2013, Brad Beckmann wrote:
-----------------------------------------------------------
This is an automatically generated e-mail. To reply, visit:
http://reviews.gem5.org/r/1849/#review4341
-----------------------------------------------------------
Ugh...another patch is going to cause us a lot of pain to merge with.
We have a lot of internal code that add experiment specific information
to the cache profiler. Those will all need to be changed quite a bit if
this patch is checked in as is.
I agree. But I believe that the patch is move towards something better
than the current state.
The current patch moves some of the simple statistics from the
CacheProfiler to the controller sim objects. However there are other,
more complicated statistics that the Cache Profiler is convenient for.
I think you should provide a concrete example for such a statistic. In my
opinion, collecting statistics in the protocol files is better as the
collection process has more visibility. The computation part can be either
hidden in the AbstractController or the CacheMemory class using the
Formula class provided in the statistics package. As an example, the
currently reported number for accesses that missed in the L2 cache in
moesi hammer protocol is incorrect.
Furthermore, this patch doesn't appear to support all the existing m5
statistics, correct? Can this patch be extended to support all m5
statistics, histograms, etc.?
You are right that the patch does not provide support for all the
structures. We would need to support the array operator [] for data
structures like histogram and vector. I will do this when I'll move more
statistics from ruby.stats to stats.txt, but I would prefer keeping the
patch as is.
Also can this patch move all the existing
stats out of CacheProfiler without actually removing the file?
While it should be possible to do so, I am not keen on it. I think it
would error prone for everyone. I would rather suggest that you maintain
the file at your end.
--
Nilay
_______________________________________________
gem5-dev mailing list
[email protected]
http://m5sim.org/mailman/listinfo/gem5-dev
