benchmarking latency to create and install instrumentation

benchmarking latency to create and install instrumentation

[William Cohen]

[-- Attachment #1: Type: text/plain, Size: 760 bytes --]

Earlier this week I added some instrumentation to the testsuite to 
record the amount of time required to do the various phases in the 
systemtap instrumentation. This information is generated for anything in 
the testsuite that uses lib/stap_run.exp or lib/stap_run2.exp.

There is a really stupid script, tests/tools/stap_metrics that extracts 
that information and can mail it off.

I have massaged the data from the FC5 nightly runs for i686 and x86_64 
into an open office spread sheet.  The i686 machine I am using is a 
relatively slow pentium III machine. The basic characteristic to note is 
that much of the latency is due to the build of the object code from the 
generated C code in phase 4.

I have attached the spread sheet to this email.

-Will

[-- Attachment #2: stap_metrics_200603030930.ods --]
[-- Type: application/vnd.oasis.opendocument.spreadsheet, Size: 14166 bytes --]

Re: benchmarking latency to create and install instrumentation

[Frank Ch. Eigler]

wcohen wrote:

> [...] information is generated for anything in the testsuite that
> uses lib/stap_run.exp or lib/stap_run2.exp. [...]

It may be just as useful to write one special .exp test cases that
extract those numbers for the limiting scenarios (say, a nearly blank
probe script, and a slow one with syscalls.*, and maybe one
intermediate "typical" one).  The .exp script could pass the times
straight into the test summary.  No need for email or spreadsheets.

- FChE