Re: testsuite and hardcoded timeouts

[William Cohen <wcohen@redhat.com>]

Quentin Barnes wrote:
> I mentioned this issue as an aside and asked about it a month ago,
> but I don't think it got a response at the time.
> 
> In porting the Systemtap testsuite to an embedded ARM platform
> (~350Mhz CPU with 64MB and running using an NFS root and swap), I
> found many of the existing hardcoded timeout parameters are way too
> short.  Several of them I had to at least triple if not increase by
> a larger factor, sometimes 6x-15x to get them to pass.
> 
> How do we want to deal with this portability problem of hardcoded
> timeouts?
> 
> 
> There are a few ways I can think of to address this:
> 
> 1) Let the hardcoded numbers stay, but up them large enough to
>   handle the slowest platform we might ever run on.  If that's
>   still not slow enough someday, up them some more when the time
>   comes.
> 
> This has the advantage of simplicity, but can greatly slow down
> suite runs on faster processors when tests do get stuck.
> 
> 2) Ban all standalone hardcoded timeouts replacing them with an
>   expression involving a multiplier and/or a constant and a
>   multiplier.
> 
> This is not the cleanest because some tests are slow due to I/O
> bandwidth or paging where others are slow due to CPU limitations.
> But it does have the advantage that if someone is having timeout
> issues, they can up the multiplier value and rerun to see if the
> problem goes away without having to edit all sorts of wrapper
> scripts and tests.
> 
> If we go with a multiplier, the multiplier could be set
> automatically by reading the cpuinfo and taking a stab at it based
> on the machine's BogoMIPS or MHz.  We'd still need a way to have a user
> straightforwardly tweak it beyond that manually.
> 
> Unfortunately, I don't understand the Systemtap testsuite framework
> yet well enough to make specific suggestions.
> 
> Thoughts?
> 
> Quentin

Hi Quentin,

I have some machines regularly downloading cvs snapshots of systemtap and 
running the tests. I have encountered the same problem, particularly on the slow 
pentium III machine. I have increased some of the timeouts as a result of this. 
However, the problem is we don't know how long some of the tests take to run. In 
addition to the processor speed the kernel/debuginfo could affect the time 
required to build/install the tests.

I don't have good solutions to this problem. However, it might be good to start 
listing the tests that are "too slow."  People running probe might be okay with 
a script taking a little time to get started, but they might not be so patient 
when it takes minutes for the script to translate and start running. Running 
them by hand with the "-v" to get information about which phases time is being 
spent would be helpful.

-Will