[Date Prev][Date Next] [Thread Prev][Thread Next] [Date Index] [Thread Index]

Re: Bug#994426: octave-sparsersb: flaky autopkgtest on ci.d.n armhf worker



On 20210916@20:39, Paul Gevers wrote:
> Hi Michele,
> 
> On 16-09-2021 19:58, Michele Martone wrote:
> > You are (well, your CI is) building librsb with support for a limited threads 
> > count, but invoking it on a machine with lots of cores, without specifying
> > the limited threads count.
> 
> I don't understand what you're saying here. Is the autopkgtest building
> a binary that's unsuitable for it's own use and that fails? That would
> be a bug in the autopkgtest. However, if you're saying that the binaries
> as build in the archive are unsuitable for hosts with many cores, than I
> don't think the package is suitable for Debian as-is. In Debian we build
> once, and expect the binary to run sanely on any machine that meets the
> minimal requirements of that architecture. Debian doesn't (and never
> has) build binaries to match the host the binary is installed onto.
> Switching build parameters based on properties of the build host is
> considered a serious bug in Debian.
> 
> > So the behavior of librsb wrt this situation (it complains) is sane --
> > the user building it shall also know on how many threads to run it (and 
> > 160 won't make any sense for the next few years)..
> 
> I claim that for the package to be suitable for Debian, if this matters,
> it needs to be determined at run time, not at build time.
> 
> I think we agree that this host is extremely powerful.
It is.

> However, Debian
> packages are expected to behave, also on those hosts.

Librsb is a package for efficient multithreaded sparse algebra.
There is no contemporary CPU I'm aware where you can run efficient 
sparse algebra with 160 cores (memory bandwidth won't cope with that).
So any savvy user uses librsb with a number of threads which is much
less than the cores count.
With dense algebra the story is different -- there using all cores makes sense.

Anyway, your points are all sound, sure.
So I'll fix it in a way the CI will pass -- thanks for making me aware of all this.

And if you've specific questions, just reiterate.

Ciao, Michele


Reply to: