- Synopsis
- Description
- Options
- Usage
- See also
- source code: bin/s370_perf_sum
s370_perf_sum [OPTIONS]... [FILE]...
s370_perf_sum reads multiple s370_perf_ana output files and allows to a create compact, one line per s370_perf test, listing summarizing and comparing the benchmark cases. The output format is, for an example with three input files and using the -rat option
File num: name ----------------------------------- #test w50-med w50-max
01: 2018-03-04_sys2.dat 303 0.78% 3.8%
02: 2018-01-09_nbk2.dat 230 2.56% 8.4%
03: 2018-01-03_rasp2b.dat 233 0.56% 9.0%
Tag Comment : tpi01 tpi02 tpi03 : t02/t01 t03/t01
T100 LR R,R : 2.09 3.56 28.98 : 1.703 13.866
T101 LA R,n : 2.76 4.10 49.77 : 1.486 18.033
T102 L R,m : 7.53 11.94 130.93 : 1.586 17.388
T103 L R,m (unal) : 8.61 15.04 138.38 : 1.747 16.072
T104 LH R,m : 9.13 15.19 142.14 : 1.664 15.568
T105 LH R,m (unal3) : 9.54 15.81 146.03 : 1.657 15.307
...
The header section lists the input files plus some statistics
- #test: number of s370_perf tests found in this file
- w50-med: median of the distribution width in %, good to judge data reproducibility
- w50-max: maximal distribution width in %
The main body has the columns
- Tag: the s370_perf test name
- Comment: text describing this test
- tpi..: instruction time from case nn
- t../t01: relative time, case nn normalized by the first case 01
See the description of -nrr, -nrx, -min, and -w50 for options which allow to display other data fields, the description of -fmis and -fsig for options to filter output lines, and the description of -k, -r, and -i for options to sort the output.
| Option | Description |
|---|---|
| -nrr | use n-rr (T100 normalized) |
| -nrx | use n-rx (T102 normalized) |
| -min | use min (minimal time) |
| -w50 | use w50 (50% distribution width) |
| -rel | show relative speed for each tag |
| -rat | show in addition ratios relative to 1st file |
| -fmis[=n] | drop test line when n values missing |
| -fsig[=n] | show test line only when differing by n percent |
| -k=key | select sort criterion |
| -r | reverse sorting order |
| -i | select instruction time lines |
| -help | print help text and quit |
Use the n-rr field instead of the cor field from the input data.
Instead of the absolute instruction time in ns the LR instruction normalized
relative time is used. This way the absolute speed of the hardware
is normalized out and systematic differences can be studied.
The output columns are labeled nrr** instead of tpi**.
See example for usage of -nrr.
Use the n-rx field instead of the cor field from the input data.
Like for -nrr, but with L instruction normalized
relative time.
The output columns are labeled nrx** instead of tpi**.
Use the min field instead of the cor field from the input data.
Useful for debugging, use the minimal instead of the median time.
Caveat is that in this case no loop overhead is subtracted.
The output columns are labeled min** instead of tpi**.
Use the w50 field instead of the cor field from the input data.
Useful for debugging, to visualize the distribution width.
The output columns are labeled w50_** instead of tpi**.
Will determine, for each test independently, the minimal value and normalize the data to this minimal value. The fastest case has per construction a value of 1., all others are >= 1., the value indicating how much slower a case is compared to the fastest one. Note that this can be combined with -nrr or -nrx.
Adds columns with ratios relative to the 1st file. This allows to have the absolute instruction times and ratios in one output file. Note that proper choice of file order is important here as the 1st one is used as normalization and thus as reference case.
The input files may contain different sets of tests, e.g. because
they contain data of different s370_perf revisions. When data is missing
for an input file ' -' is written into the summary output.
If -fmis=n is specified tests lines are dropped when n or more values
are missing. A -fmis=1, which can be abbreviated to -fmis, will drop
all tests which are missing in one or more input files.
Show test line only when differing by at least n percent. This option is
very helpful when comparing two test campaigns on the same system to verify
the reproducibility. With -fsig=1, which can be abbreviated to -fsig,
only tests where one case differs by at least 1 percent are listed.
Selects the sorting criterion, if no -k option is given -k=tag is assumed.
Supported key values are
- tag - the s370_perf test id (Tnnn) or calculation id (Dnnn). In most cases this simply reproduces the input file because normal s370_perf_ana output is sorted by test id.
- ins - selects only tests which give the timing of a single instruction, and prints these tests ordered by instruction mnemonic.
- vN - sort by value of N'th file, e.g. v1,v2,...
- rN - sort by ratio for N'th file, e.g. r2,...
See example for usage of -k=r2.
The default sorting order is ascending, this option reverses the sorting order
to descending. Supported only for value (like -k=v1) and ratio sorts
(line -k=r2).
Selects only tests which give the timing of a single instruction. All composite and auxiliary tests are removed.
Print help text and quit.
To list the instruction times of test cases plus the time rations relative to the first case use
s370_perf_sum -rat <file1> <file2> ...
An example output is shown in the description section.
When comparing data from different systems it is interesting to check
whether the different instructions all show the same speed ratio or
whether there are significant differences. Relative speed differences
can be seen best when the absolute speed of the underlying hardware
normalized out by dividing all instruction times by the time for LR
with the -nrr option, like in
s370_perf_sum -nrr -rat <file1> <file2> ...
The value listed for T100 is now by construction 1.00, the ratio in the last
column(s) is now a double ratio
( tpi(test,filex) / tpi(T100,filex) ) / ( tpi(test,file1) / tpi(T100,file1) )
which equivalent to
( tpi(test,filex) / tpi(test,file1) ) / ( tpi(T100,filex) / tpi(T100,file1) )
which indicates the variation of the relative speed difference. A practical example is given in the narrative for dataset 2018-01-03_rasp2b.
The body of the output looks like (same data as shown in description section)
Tag Comment : nrr01 nrr02 nrr03 : t02/t01 t03/t01
T100 LR R,R : 1.00 1.00 1.00 : 1.000 1.000
T101 LA R,n : 1.32 1.15 1.72 : 0.871 1.303
T102 L R,m : 3.59 3.35 4.52 : 0.933 1.259
T103 L R,m (unal) : 4.11 4.22 4.77 : 1.027 1.161
T104 LH R,m : 4.36 4.27 4.90 : 0.979 1.124
T105 LH R,m (unal3) : 4.56 4.44 5.04 : 0.974 1.105
T106 LTR R,R : 1.68 1.47 1.31 : 0.875 0.780
T107 LCR R,R : 1.73 1.46 1.50 : 0.844 0.867
...
This analysis can also be done with a normalization on the L
instruction time with the -nrx option.
For a quick overview of what is faster or slower it helps to sort the output by the speed ratio with the -k option, like in
s370_perf_sum -rat -k r2 <file1> <file2> ...
which gives an output like
File num: name ----------------------------------- #test w50-med w50-max
01: 2018-03-31_sys2.dat 330 0.89% 4.5%
02: 2018-04-02_rasp2b.dat 330 0.54% 15.8%
Tag Comment : tpi01 tpi02 : t02/t01
T156 MVC m,m (250c,over1) : 351.62 1140.12 : 3.242
T276 CLC m,m (250c,eq) : 165.90 667.73 : 4.025
T274 CLC m,m (100c,eq) : 83.62 431.75 : 5.163
D610 EX R,i (bare, via TM) : 21.75 133.29 : 6.128
D611 EX R,i (bare, via XI) : 20.85 132.73 : 6.366
...
D215 DR R,R : 14.72 498.06 : 33.836
T600 STCK m : 65.48 2838.00 : 43.341
T547 DD R,m : 156.37 7336.85 : 46.920
T546 DDR R,R : 152.40 7254.83 : 47.604
T401 CVD R,m : 36.04 6177.14 : 171.397
Using in addition the -nrr option helps to see the relative trends better, and the -i option helps to focus in the instruction timings, like in
s370_perf_sum -nrr -rat -i -k r2 <file1> <file2> ...
which gives an output like (same files as above)
Tag Comment : nrr01 nrr02 : t02/t01
T156 MVC m,m (250c,over1) : 168.33 39.42 : 0.234
T276 CLC m,m (250c,eq) : 79.42 23.09 : 0.291
T274 CLC m,m (100c,eq) : 40.03 14.93 : 0.373
D610 EX R,i (bare, via TM) : 10.41 4.61 : 0.443
D611 EX R,i (bare, via XI) : 9.98 4.59 : 0.460
T304 BR R : 3.21 1.67 : 0.520
...
T600 STCK m : 31.35 98.12 : 3.130
T547 DD R,m : 74.86 253.67 : 3.389
T546 DDR R,R : 72.95 250.83 : 3.438
T401 CVD R,m : 17.25 213.57 : 12.381
- s370_perf - IBM System/370 Instruction Timing Benchmark
- s370_perf_ana - analyze s370_perf data
- s370_perf_sort - generate a sorted s370_perf data listing
- s370_perf_mark - derive MIPS ratings from s370_perf data