[parsec-users] suitable L2$ size to capture 1st order temporal locality

Muhammad abid Mughal mabidm_pieas at yahoo.com
Fri Sep 3 16:45:15 EDT 2010


Thank you very much.



________________________________
From: Christian A Bienia <cbienia at CS.Princeton.EDU>
To: PARSEC Users <parsec-users at lists.cs.princeton.edu>
Sent: Saturday, September 4, 2010 2:45:57
Subject: Re: [parsec-users] suitable L2$ size to capture 1st order temporal 
locality

Hi Muhammad,

There is no clear distinction between first-level and second-level working set 
sizes. The sizes I used in the PARSEC characterization paper are more or less 
arbitrary. How much of a cache is a 'good' size depends on the application - 
what is enough for one program might be completely laughable for another. It's 
best if you use the graphs I provide in the characterization paper to make a 
rough size estimate and use something that makes sense for your research 
project.

Best,
Chris


----- Ursprüngliche Mail -----
Von: "Muhammad abid Mughal" <mabidm_pieas at yahoo.com>
An: "Parsec group" <parsec-users at lists.cs.princeton.edu>
Gesendet: Freitag, 3. September 2010 17:06:32 GMT +01:00 
Amsterdam/Berlin/Bern/Rom/Stockholm/Wien
Betreff: [parsec-users] suitable L2$ size to capture 1st order temporal locality




hi , 
Hope you GUYS doing good.I am running some simulations with 8-node distributed 
memory system,with different L2$ size,to measure suitable L2$ size which can 
capture 1st order temporal locality (working set 1) .I am using "off-chip read 
misses per 1000 instructions" to measure it.In your paper titled "The PARSEC 
Benchmark Suite: Characterization and Architectural Implications" you show 
working set1 and working set2 for "simlarge" input set.I am using "simsmall" 
dataset so definitely these working sets would be smaller.But how do i know that 
this L2$ size has captured 1st order temporal locality or 2n order(i.e. working 
set2).I must know the data structure and its size being used by parsec 
workload.. Is there anyway to get such kind of information without going through 
code?(because i am not a programmer). 

For "simsmall" I think 8MB L2$ is enough to capture 1st order temporal locality 
of PARSEC . 



My system parameters and results are: 
************************************************************************* 

Input: "simsmall" 
Node : L1 I/D cache: 64KB each;L2 Cache: 1/2/4/8 MB;one processor/chip 
System: 8-node distributed shared memory system 
Total memory: 4GB 
Task : Measuring off-chip read misses per 1000 instructions to know the suitable 
size of L2$. 

profilingfiles1MBL2: Profiling files related to L2$=1MB 
profilingfiles2MBL2: Profiling files related to L2$=2MB 
profilingfiles4MBL2: Profiling files related to L2$=4MB 
profilingfiles8MBL2: Profiling files related to L2$=8MB 


For each workload i just varied the L2$ size and measured "off-chip read misses 
per 1000 instructions". 

The results are given below 


./profilingfiles1MBL2/freqmine-simsmall.profile:Off-chip_readmisses_per_1000_insts
 0.0260005 

./profilingfiles2MBL2/freqmine-simsmall.profile:Off-chip_readmisses_per_1000_insts
 0.0258792 

./profilingfiles4MBL2/freqmine-simsmall.profile:Off-chip_readmisses_per_1000_insts
 0.0258433 

/profilingfiles8MBL2/freqmine-simsmall.profile:Off-chip_readmisses_per_1000_insts
 0.0258328 





./profilingfiles1MBL2/bodytrack-simsmall.profile:Off-chip_readmisses_per_1000_insts
 0.0752952 

./profilingfiles2MBL2/bodytrack-simsmall.profile:Off-chip_readmisses_per_1000_insts
 0.0762126 

./profilingfiles4MBL2/bodytrack-simsmall.profile:Off-chip_readmisses_per_1000_insts
 0.075893 

./profilingfiles8MBL2/bodytrack-simsmall.profile:Off-chip_readmisses_per_1000_insts
 0.0750404 





./profilingfiles1MBL2/canneal-simsmall.profile:Off-chip_readmisses_per_1000_insts
 2.3342 

./profilingfiles2MBL2/canneal-simsmall.profile:Off-chip_readmisses_per_1000_insts
 2.29017 

./profilingfiles4MBL2/canneal-simsmall.profile:Off-chip_readmisses_per_1000_insts
 2.28839 

./profilingfiles8MBL2/canneal-simsmall.profile:Off-chip_readmisses_per_1000_insts
 2.22812 





./profilingfiles1MBL2/vips-simsmall.profile:Off-chip_readmisses_per_1000_insts 
0.715399 

./profilingfiles2MBL2/vips-simsmall.profile:Off-chip_readmisses_per_1000_insts 
0.676205 

./profilingfiles4MBL2/vips-simsmall.profile:Off-chip_readmisses_per_1000_insts 
0.697383 

./profilingfiles8MBL2/vips-simsmall.profile:Off-chip_readmisses_per_1000_insts 
0.64845 





./profilingfiles1MBL2/fluidanimate-simsmall.profile:Off-chip_readmisses_per_1000_insts
 0.135831 

./profilingfiles2MBL2/fluidanimate-simsmall.profile:Off-chip_readmisses_per_1000_insts
 0.135771 

./profilingfiles4MBL2/fluidanimate-simsmall.profile:Off-chip_readmisses_per_1000_insts
 0.135985 

./profilingfiles8MBL2/fluidanimate-simsmall.profile:Off-chip_readmisses_per_1000_insts
 0.135115 





./profilingfiles1MBL2/dedup-simsmall.profile:Off-chip_readmisses_per_1000_insts 
0.0616364 

./profilingfiles2MBL2/dedup-simsmall.profile:Off-chip_readmisses_per_1000_insts 
0.0693683 

./profilingfiles4MBL2/dedup-simsmall.profile:Off-chip_readmisses_per_1000_insts 
0.338068 

./profilingfiles8MBL2/dedup-simsmall.profile:Off-chip_readmisses_per_1000_insts 
0.0652708 





./profilingfiles1MBL2/blackscholes-simsmall.profile:Off-chip_readmisses_per_1000_insts
 0.846945 

./profilingfiles2MBL2/blackscholes-simsmall.profile:Off-chip_readmisses_per_1000_insts
 0.814208 

./profilingfiles4MBL2/blackscholes-simsmall.profile:Off-chip_readmisses_per_1000_insts
 0.791456 

./profilingfiles8MBL2/blackscholes-simsmall.profile:Off-chip_readmisses_per_1000_insts
 0.818974 





./profilingfiles1MBL2/facesim-simsmall.profile:Off-chip_readmisses_per_1000_insts
 0.347814 

./profilingfiles2MBL2/facesim-simsmall.profile:Off-chip_readmisses_per_1000_insts
 0.348019 

./profilingfiles4MBL2/facesim-simsmall.profile:Off-chip_readmisses_per_1000_insts
 0.294693 

./profilingfiles8MBL2/facesim-simsmall.profile:Off-chip_readmisses_per_1000_insts
 0.292923 





./profilingfiles1MBL2/x264-simsmall.profile:Off-chip_readmisses_per_1000_insts 
0.0280423 

./profilingfiles2MBL2/x264-simsmall.profile:Off-chip_readmisses_per_1000_insts 
0.0284851 

./profilingfiles4MBL2/x264-simsmall.profile:Off-chip_readmisses_per_1000_insts 
0.0284456 

./profilingfiles8MBL2/x264-simsmall.profile:Off-chip_readmisses_per_1000_insts 
0.0288943 





./profilingfiles1MBL2/ferret-simsmall.profile:Off-chip_readmisses_per_1000_insts 
0.268492 

./profilingfiles2MBL2/ferret-simsmall.profile:Off-chip_readmisses_per_1000_insts 
0.271471 

./profilingfiles4MBL2/ferret-simsmall.profile:Off-chip_readmisses_per_1000_insts 
0.267186 

./profilingfiles8MBL2/ferret-simsmall.profile:Off-chip_readmisses_per_1000_insts 
0.210023 





./profilingfiles1MBL2/streamcluster-simsmall.profile:Off-chip_readmisses_per_1000_insts
 7.084 

./profilingfiles2MBL2/streamcluster-simsmall.profile:Off-chip_readmisses_per_1000_insts
 6.86997 

./profilingfiles4MBL2/streamcluster-simsmall.profile:Off-chip_readmisses_per_1000_insts
 6.68449 

./profilingfiles8MBL2/streamcluster-simsmall.profile:Off-chip_readmisses_per_1000_insts
 6.48975 





./profilingfiles1MBL2/swaptions-simsmall.profile:Off-chip_readmisses_per_1000_insts
 1.56026 

./profilingfiles2MBL2/swaptions-simsmall.profile:Off-chip_readmisses_per_1000_insts
 1.29674 

./profilingfiles4MBL2/swaptions-simsmall.profile:Off-chip_readmisses_per_1000_insts
 1.60086 

./profilingfiles8MBL2/swaptions-simsmall.profile:Off-chip_readmisses_per_1000_insts
 1.72674 





./profilingfiles1MBL2/raytrace-simsmall.profile:Off-chip_readmisses_per_1000_insts
 0.0721749 

./profilingfiles2MBL2/raytrace-simsmall.profile:Off-chip_readmisses_per_1000_insts
 0.0763548 

./profilingfiles4MBL2/raytrace-simsmall.profile:Off-chip_readmisses_per_1000_insts
 0.0768016 

./profilingfiles8MBL2/raytrace-simsmall.profile:Off-chip_readmisses_per_1000_insts
 0.0746796 



****************************************************************************** 


As we can see , there is a very small change in 
"Off-chip_readmisses_per_1000_insts" when we vary L2$ size from 1MB to 8MB. 





Hoping someone will give a prompt response. 




Regards, 
Muhammad abid 


_______________________________________________
parsec-users mailing list
parsec-users at lists.cs.princeton.edu
https://lists.cs.princeton.edu/mailman/listinfo/parsec-users
_______________________________________________
parsec-users mailing list
parsec-users at lists.cs.princeton.edu
https://lists.cs.princeton.edu/mailman/listinfo/parsec-users


-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.cs.princeton.edu/pipermail/parsec-users/attachments/20100903/494e45d1/attachment.htm>


More information about the parsec-users mailing list