Discussion:
[gem5-dev] Review Request 1667: sim: simulate with multiple event queues
(too old to reply)
Nilay
2013-11-17 19:30:57 UTC
Permalink
I have attached profiling results from two runs. While overall we do not
see much difference, but it seems some regressions are gaining and some
are losing.

Here is a summary of the two runs. The data is the difference between
the time taken for the run with the multi-threading patch applied and
the time taken for the run without it.


Run 1
-----

All regression tests:
---------------------
Maximum Gain in Time (s) 305.37
Average Gain in Time (s) -33.2935971223

Maximum % 125
Average % -1.0315445599

Only long regression tests:
---------------------------
Maximum (long) 305.37
Average (long) -66.0515714286

Maximum % (long) 15.7268292683
Average % (long) -0.9671363525


Run 2
-----

All regression tests:
---------------------
Maximum Gain in Time (s) 552.08
Average Gain in Time -0.3939568345

Maximum % 60
Average % -2.1395937762

Only long regression tests:
---------------------------
Maximum (long) 552.08
Average (long) -1.534

Maximum % (long) 17.0595645132
Average % (long) -0.4627593543


--
Nilay
Hi Nilay,
I haven't tried OS X. The results I reported are RHE5 and Ubuntu 10.04,
both with gcc >= 4.7.
Andreas
Hi Nilay,
Unfortunately I haven't found the time to dig into this yet.
I only have the output from last nights regression, and thus not with
and
without your patch. I'll take a snapshot and compare, although I
suspect
the individual runs vary quite a bit.
Have you attempted any further profiling?
No, I have not. Are you running on OS X?
--
Nilay
Nilay Vaish
2013-11-17 22:31:27 UTC
Permalink
Seems like the attached spreadsheet was removed by the mail server. Here
is link to a Google Doc:
https://docs.google.com/spreadsheet/ccc?key=0AmlHkQaBlMvBdEhBU28zYThPSHJybjlWQzdZNUpBbGc&usp=sharing

You should be able to access it even without a Google account.

--
Nilay
I have attached profiling results from two runs. While overall we do not see
much difference, but it seems some regressions are gaining and some are
losing.
Here is a summary of the two runs. The data is the difference between the
time taken for the run with the multi-threading patch applied and the time
taken for the run without it.
Run 1
-----
---------------------
Maximum Gain in Time (s) 305.37
Average Gain in Time (s) -33.2935971223
Maximum % 125
Average % -1.0315445599
---------------------------
Maximum (long) 305.37
Average (long) -66.0515714286
Maximum % (long) 15.7268292683
Average % (long) -0.9671363525
Run 2
-----
---------------------
Maximum Gain in Time (s) 552.08
Average Gain in Time -0.3939568345
Maximum % 60
Average % -2.1395937762
---------------------------
Maximum (long) 552.08
Average (long) -1.534
Maximum % (long) 17.0595645132
Average % (long) -0.4627593543
--
Nilay
Hi Nilay,
I haven't tried OS X. The results I reported are RHE5 and Ubuntu 10.04,
both with gcc >= 4.7.
Andreas
Hi Nilay,
Unfortunately I haven't found the time to dig into this yet.
I only have the output from last nights regression, and thus not with
and
without your patch. I'll take a snapshot and compare, although I
suspect
the individual runs vary quite a bit.
Have you attempted any further profiling?
No, I have not. Are you running on OS X?
--
Nilay
Nilay Vaish
2013-11-20 12:18:10 UTC
Permalink
I had another run with the multithread patch applied and the time required
remains about the same. Unless anyone objects, I'll commit the patch
tomorrow.

--
Nilay
Seems like the attached spreadsheet was removed by the mail server. Here is
https://docs.google.com/spreadsheet/ccc?key=0AmlHkQaBlMvBdEhBU28zYThPSHJybjlWQzdZNUpBbGc&usp=sharing
You should be able to access it even without a Google account.
--
Nilay
Andreas Hansson
2013-11-20 14:28:45 UTC
Permalink
Hi Nilay,

I hate to be a pain, but I¹m still wondering what caused the 15-20% slow
down for the regressions on our side. Could you hold off until Monday next
week?

Thanks,

Andreas
Post by Nilay Vaish
I had another run with the multithread patch applied and the time required
remains about the same. Unless anyone objects, I'll commit the patch
tomorrow.
--
Nilay
Seems like the attached spreadsheet was removed by the mail server. Here is
https://docs.google.com/spreadsheet/ccc?key=0AmlHkQaBlMvBdEhBU28zYThPSHJy
bjlWQzdZNUpBbGc&usp=sharing
You should be able to access it even without a Google account.
--
Nilay
_______________________________________________
gem5-dev mailing list
http://m5sim.org/mailman/listinfo/gem5-dev
-- IMPORTANT NOTICE: The contents of this email and any attachments are confidential and may also be privileged. If you are not the intended recipient, please notify the sender immediately and do not disclose the contents to any other person, use it for any purpose, or store or copy the information in any medium. Thank you.

ARM Limited, Registered office 110 Fulbourn Road, Cambridge CB1 9NJ, Registered in England & Wales, Company No: 2557590
ARM Holdings plc, Registered office 110 Fulbourn Road, Cambridge CB1 9NJ, Registered in England & Wales, Company No: 2548782
Nilay Vaish
2013-11-20 14:35:06 UTC
Permalink
OK.
Hi Nilay,
I hate to be a pain, but I¹m still wondering what caused the 15-20% slow
down for the regressions on our side. Could you hold off until Monday next
week?
Thanks,
Andreas
Post by Nilay Vaish
I had another run with the multithread patch applied and the time required
remains about the same. Unless anyone objects, I'll commit the patch
tomorrow.
--
Nilay
Andreas Hansson
2013-11-25 15:28:36 UTC
Permalink
Nilay Vaish
2013-11-25 16:53:36 UTC
Permalink
Thanks Andreas. I am going to push the patch soon.

--
Nilay
Hi Nilay,
Thanks for your patience. Here is the verdict after averaging over two
days both with and without your patch. I have compared the host_seconds
with and without your patch for all the regressions, and here list those
that require at least 10 seconds to run, and change more than 10%. For
each of the included benchmarks I have printed the name, the ratio of
before/after, and the absolute host_seconds before after. Not a single o3
run pops up, and no switcheroos. Instead it is the simple CPUs that stand
out.
X86 seems to be ~15% slower, along with SPARC, and ARM ~15% faster (any
idea why?). I don¡¯t think the differences are significant enough to hold
off on the patch.
Thanks,
Andreas
/X86/tests/opt/long/se/60.bzip2/x86/linux/simple-atomic
0.846341761811
2144.52 2533.87
/X86/tests/opt/long/se/60.bzip2/x86/linux/simple-timing
0.880894867374
3821.78 4338.52
/X86/tests/opt/long/se/20.parser/x86/linux/simple-atomic
0.85584925128
630.41 736.59
/X86/tests/opt/long/se/20.parser/x86/linux/simple-timing
0.898765239746
1153.7 1283.65
/X86/tests/opt/long/se/10.mcf/x86/linux/simple-atomic
0.86398415084
126.47 146.38
/X86/tests/opt/long/se/10.mcf/x86/linux/simple-timing
0.841928537538
230.68 273.99
/X86/tests/opt/long/se/70.twolf/x86/linux/simple-atomic
0.851222651223
99.21 116.55
/X86/tests/opt/long/se/70.twolf/x86/linux/simple-timing
0.889776357827
172.67 194.06
/ARM/tests/opt/long/se/60.bzip2/arm/linux/simple-timing
1.12286217903
1649.9 1469.37
/ARM/tests/opt/quick/fs/10.linux-boot/arm/linux/realview-simple-atomic-dual
1.1978988517
49.03 40.93
/ARM/tests/opt/quick/fs/10.linux-boot/arm/linux/realview-switcheroo-atomic
1.17865429234
55.88 47.41
/ARM/tests/opt/quick/fs/10.linux-boot/arm/linux/realview-simple-atomic
1.17839607201
50.4 42.77
/SPARC/tests/opt/long/se/10.mcf/sparc/linux/simple-atomic
0.770113942768
88.54 114.97
/SPARC/tests/opt/long/se/10.mcf/sparc/linux/simple-timing
0.866449660737
188.99 218.12
/SPARC/tests/opt/long/se/70.twolf/sparc/linux/simple-atomic
0.766434648105
69.37 90.51
/SPARC/tests/opt/long/se/70.twolf/sparc/linux/simple-timing
0.799702080119
150.32 187.97
/SPARC/tests/opt/long/se/50.vortex/sparc/linux/simple-atomic
0.763137377203
48.94 64.13
/SPARC/tests/opt/long/se/50.vortex/sparc/linux/simple-timing
0.845123980036
106.68 126.23
OK.
Hi Nilay,
I hate to be a pain, but I©öm still wondering what caused the 15-20% slow
down for the regressions on our side. Could you hold off until Monday next
week?
Thanks,
Andreas
Post by Nilay Vaish
I had another run with the multithread patch applied and the time required
remains about the same. Unless anyone objects, I'll commit the patch
tomorrow.
--
Nilay
-- IMPORTANT NOTICE: The contents of this email and any attachments are confidential and may also be privileged. If you are not the intended recipient, please notify the sender immediately and do not disclose the contents to any other person, use it for any purpose, or store or copy the information in any medium. Thank you.
ARM Limited, Registered office 110 Fulbourn Road, Cambridge CB1 9NJ, Registered in England & Wales, Company No: 2557590
ARM Holdings plc, Registered office 110 Fulbourn Road, Cambridge CB1 9NJ, Registered in England & Wales, Company No: 2548782
_______________________________________________
gem5-dev mailing list
http://m5sim.org/mailman/listinfo/gem5-dev
Loading...