I stumbled some time ago on a page about Hercules instruction timing, didn't bookmark it properly, and failed to find it again.

I'm fully aware of

Despite all these provisos I wrote a small synthetic benchmark to determine the 370 instruction timing. It covers most non-privileged instructions in a total of 124 individual tests. The assembler source and a result file are uploaded to Yahoo! Group turnkey-mvs under perf_asm.zip.

The job perf_asm.jcl runs about 2 minutes on a modern x86 CPU. I've executed the job on a Turnkey 4- system under Linux. So the results reflect the performance of the Hercules version bundles with Juergen Winkelmann's TK4- distribution.

The results files perf_asm.txt list the test tag name 'Tnnn', a short description of the instruction tested.

The timing is actually measured with STCK, which is a 'time of day' clock, The tests were done on an otherwise idle MVS system, and on an otherwise idle host system. The job perf_asm.jcl was executed 29 times, the value shown in column 'med' are the median values of these 29 runs to have an estimate robust against outliers.

There are many obvious and expected trends and relations, but also a few surprises (at least for me).

I consider this V0.1 of this code, it needs further refinement and cross-checking. I'd very much appreciate

2018-01-06: The code was renamed to s370-perf and is now the central element of a a separate GitHub project wfjm/s370-perf.
For original posting to Yahoo! Group - turnkey-MVS see topic 10720. Dead link since 2020-12-15: Yahoo! Groups was discontinued by Verizon.