unreasonable multiple Whetstone results about byte-unixbench

[GoogleCodeExporter]

sorry, the title should be "unreasonable multiple-core Whetstone results"

Original comment by [email protected] on 5 Dec 2014 at 2:40

[GoogleCodeExporter]

I assume this has got to do with how the measurement is done. Just raised the 
same issue. I changed the flags passed to compiler and it fixed the problem. 
Could you kindly try the same?

Instead of -DUNIX , kindly use -DGTODay so that it uses wall clock time instead 
of process cpu time.

Thanks.

Original comment by [email protected] on 30 Mar 2015 at 1:11

[GoogleCodeExporter]

tested in OS X, 4C8T, (not idle). looks it still uses process cpu time. 
1: 1004.1
2: 1951.8
4: 3887.8
8: 6794.8

Original comment by [email protected] on 30 Mar 2015 at 2:41

[GoogleCodeExporter]

I tried now and I see it is working fine.

DUNIX:

Benchmark Run: Mon Mar 30 2015 21:48:04 - 21:48:19
4 CPUs in system; running 1 parallel copy of tests

Double-Precision Whetstone                     4203.4 MWIPS (9.7 s, 1 samples)

System Benchmarks Partial Index              BASELINE       RESULT    INDEX
Double-Precision Whetstone                       55.0       4203.4    764.3
                                                                   ========
System Benchmarks Index Score (Partial Only)                          764.3

------------------------------------------------------------------------
Benchmark Run: Mon Mar 30 2015 21:48:19 - 21:48:36
4 CPUs in system; running 4 parallel copies of tests

Double-Precision Whetstone                    13725.4 MWIPS (9.9 s, 1 samples)

System Benchmarks Partial Index              BASELINE       RESULT    INDEX
Double-Precision Whetstone                       55.0      13725.4   2495.5
                                                                   ========
System Benchmarks Index Score (Partial Only)                         2495.5

------------------------------------------------------------------------
Benchmark Run: Mon Mar 30 2015 21:48:36 - 21:49:05
4 CPUs in system; running 8 parallel copies of tests

Double-Precision Whetstone                    27433.0 MWIPS (9.9 s, 1 samples)

System Benchmarks Partial Index              BASELINE       RESULT    INDEX
Double-Precision Whetstone                       55.0      27433.0   4987.8
                                                                   ========
System Benchmarks Index Score (Partial Only)                         4987.8


DGTODay:

Benchmark Run: Mon Mar 30 2015 21:51:08 - 21:51:24
4 CPUs in system; running 1 parallel copy of tests

Double-Precision Whetstone                     4293.7 MWIPS (10.0 s, 1 samples)

System Benchmarks Partial Index              BASELINE       RESULT    INDEX
Double-Precision Whetstone                       55.0       4293.7    780.7
                                                                   ========
System Benchmarks Index Score (Partial Only)                          780.7

------------------------------------------------------------------------
Benchmark Run: Mon Mar 30 2015 21:51:24 - 21:51:40
4 CPUs in system; running 4 parallel copies of tests

Double-Precision Whetstone                    13406.6 MWIPS (9.9 s, 1 samples)

System Benchmarks Partial Index              BASELINE       RESULT    INDEX
Double-Precision Whetstone                       55.0      13406.6   2437.6
                                                                   ========
System Benchmarks Index Score (Partial Only)                         2437.6

------------------------------------------------------------------------
Benchmark Run: Mon Mar 30 2015 21:51:40 - 21:51:55
4 CPUs in system; running 8 parallel copies of tests

Double-Precision Whetstone                    14167.5 MWIPS (9.3 s, 1 samples)

System Benchmarks Partial Index              BASELINE       RESULT    INDEX
Double-Precision Whetstone                       55.0      14167.5   2575.9
                                                                   ========
System Benchmarks Index Score (Partial Only)                         2575.9

Original comment by [email protected] on 30 Mar 2015 at 4:23

[GoogleCodeExporter]

Since you mentioned 8 threads, could you try with 16 copies and with both the 
timing options to see how the numbers look like?

Original comment by [email protected] on 30 Mar 2015 at 4:23

[GoogleCodeExporter]

Forgot to mention additional information. My laptop had 2C4T.. so.. 4T should 
be the max for any throughput scaling..

Original comment by [email protected] on 30 Mar 2015 at 4:25

[GoogleCodeExporter]

This works for me. I made tests using Arch Linux on:

Intel(R) Xeon(R) CPU W3530 (4 cores, HT)
ARM Cortex A8 (one core)
ARM Cortex A7 (dual core)

I don't have any Mac to test, maybe you may try -DMAC ? You can see the options 
at src/whets.c at "Timer options".

Thanks.

Original comment by [email protected] on 30 Mar 2015 at 4:30

[GoogleCodeExporter]

Mac OS also should support gettimeofday( ). Could you kindly try increasing the 
number of copies to be beyond the number of cpu threads and test the same?

Original comment by [email protected] on 30 Mar 2015 at 5:01

[GoogleCodeExporter]

[deleted comment]

[GoogleCodeExporter]

[deleted comment]

[GoogleCodeExporter]

tried -DMAC, and -DMAC_TMgr

-DMAC_TMgr won't even compile, "Timer.h"  not found, maybe that's for older PPC 
Macs.

result of -DMAC on 4C8T OS X (problem seems not solved, and new problem 
introduced.)

Original comment by [email protected] on 31 Mar 2015 at 4:50

Attachments:

• 20150331.Unixbench.5.1.3.DMAC.txt

[GoogleCodeExporter]

and  tested -DGTODay in 
Linux ... 3.16.0-4-686-pae #1 SMP Debian 3.16.7-ckt4-3 (2015-02-03) i686 
GNU/Linux
(already patched 'volatile unsigned long iter;')
got this error:
1 x Arithoh  1 2 3Can't take log of -2.93709e+07 at ./Run line 935.

is the timing done by Perl script 'Run' ?

Original comment by [email protected] on 31 Mar 2015 at 4:55

[GoogleCodeExporter]

Could you kindly clarify the following? I assume the initial issue reported was 
with whetstone. Did you try -DGTODay with whetstone? What was the results with 
multiple copies ( > 8 copies ).

arith.c uses a different method for measuring time. From your previous results 
arithmetic is working properly.

Original comment by [email protected] on 31 Mar 2015 at 5:46

[GoogleCodeExporter]

hmm, looks like solved.

though the score of >4 copies is much higher  than what I expected. 

Thanks!

Original comment by [email protected] on 31 Mar 2015 at 9:45

Attachments:

• 20150331.Unixbench.5.1.3.DMAC.txt

[GoogleCodeExporter]

 -DGTODay also works well in linux. 

Original comment by [email protected] on 1 Apr 2015 at 4:42

[gstrauss]

A comment in the code suggests that rusage() is preferred over gettimeofday(), but various github issues report that wall clock time of gettimeofday() is preferred.

Suggested solution is to modify Makefile
$(PROGDIR)/whetstone-double: CFLAGS += -DDP -DUNIX -DUNIXBENCH
to replace -DUNIX with -DGTODay

[gstrauss]

#15, #18, #26 all seem to be the same issue.
@kdlucas: simple fix suggested above. Would you prefer a new pull request?

[kdlucas]

Yes, can we get a new pull request for this issue?