Status of armhf-linux and powerpc64le-linux

[Ludovic Courtès]

Moin!

Mathieu Othacehe <othacehe@gnu.org> skribis:

>>      - armhf-linux is disabled on ci.guix due to improper offloading
>>        setup (probably along the lines of
>>        <https://issues.guix.gnu.org/53463>).  Should we try and reenable
>>        it, or should we drop it?
>>
>>      - powerpc64le-linux is disabled on ci.guix since today
>>        (maintenance.git commit
>>        d641115e20973731555b586985fa81fbe293aeca).  However it did work
>>        until recently and we have one machine to offload to.  Should we
>>        fix it or drop it?  Mathieu?
>
> Yeah, we only have a single machine to offload to and each time it is
> not reachable, the "guix" specification fails on Cuirass.

How frequently does that machine become unreachable?

Its uptime right now is “only” 51 days, but it seems to have been
reliably building things so far (surprisingly so!).

> That's because we need to offload to a powerpc64le-linux machine in
> order to evaluate the guix derivation for that specific architecture
> (that's true for all the other architectures).

Maybe we should arrange to be more resilient to transient build machine
outage.

For that we need redundancy; we have it for ARM, but not for POWER9.  A
simple way to get redundancy today would be to set up transparent
emulation for POWER9 on one of the x86_64 boxes.  That’ll be
inefficient, but that’ll let Cuirass survive transient failures of that
one POWER9 box.

WDYT?

Longer-term, people interested in POWER9 should look into:

  • Purchasing, setting up, and hosting POWER9 hardware (funds held at
    the FSF are probably sufficient for that!).

  • And/or: getting in touch with companies who could sponsor us by
    providing hardware (the AArch64 port was started thanks to a
    donation by ARM).

In Cuirass, we should arrange to support partial evaluations or
per-system evaluations so that a single missing offload machine doesn’t
cause the whole evaluation to fail.

> Given the lack of workers for powerpc64le-linux I think we should drop
> it.

We can do that, but I find embarrassing to drop the architecture after
all the work people have put it “just” because of infrastructure issues.

> Regarding armhf-linux we can in theory rely on the overdrives but we
> are already struggling on aarch64-linux, we I think we should also
> drop it for now.

In theory, ci.guix has at least 3 Honeycombs (2 are currently offline)
and 2 Overdrives, so it’s not that bad, and they don’t seem to be all
that busy.

So you’re right in a way, but at the same time this seems to be an
infrastructure issue.

> Focusing on x86_64-linux, i686-linux and aarch64-linux for this release
> seems more pragmatic.

That’s radical, but maybe that’s the most reasonable option.

How about a plan like this: until next Thursday, we try to address the
infrastructure issues discussed above to estimate feasibility.  Then we
decide on the way forward.  WDYT?

If we end up dropping architectures, we’ll have to:

  1. Update the documentation (and eventually the web site).

  2. Offer a clear plan as to what it would take to reinstate those
     architectures, and probably define clear criteria for architecture
     support going forward.

Thanks,
Ludo’.