Re: Buying new Honeycombs (AArch64)

[Christopher Baines]

Ludovic Courtès <ludo@gnu.org> writes:

> Christopher Baines <mail@cbaines.net> skribis:
>
>> Ludovic Courtès <ludo@gnu.org> writes:
>>
>>> I’d like us to buy 4 new Honeycomb machines (AArch64) for use as build
>>> machines on ci.guix.gnu.org, which is really short on Arm build machines
>>> (at most 4 machines, often less depending on outage).
>
> [...]
>
>> The 2 Honeycomb machines (hatysa and hamal) running behind the bordeaux
>> build farm seem to keep up reasonably well with master, patches and
>> branches, so I'm doubtful there's a need for more machines.
>
> There are inefficiencies in Cuirass (I gather that’s what you’re
> alluding to), but it’s much better than a year ago, and more
> importantly, that alone cannot explain the low AArch64 build
> availability on ci.guix compared to POWER for example.
>
> The other issue: we’ve been down to one Honeycomb and one OverDrive for
> a few weeks already.  Redundancy would help.
>
> Last, there’s the problem that we’d like to test branches more quickly
> that is currently the case, where Arm is the limiting factor.
>
> I think that justifies more machines.

I agree that more machines could be helpful, but I'm still unsure
whether they'll help with ci.guix ARM builds.

Is there any data for CPU utilisation/load for the currently running
machines? If there is and it suggests the machine is busy, then at least
that would suggest that more machines might be put to work.

>> hatysa and hamal do have much more RAM than the Honeycomb machines
>> behind CI, so I think looking to upgrade the RAM on each of the existing
>> machines (either to 32G or 64G) would make more sense to do first. That
>> should allow making much better use of the 3 machines already available.
>
> I just checked and grunewald has 32G while overdrive1 has 8G.
>
> I think it’d be hard to upgrade the RAM as we rely on someone with
> access to the MDC data center.

As I realised when I checked afterwards, the HoneyComb machines do
actually have reasonable amounts of RAM.

>> If that is done and there's still a need for more machines, given
>> there's currently no concrete plans to expand the bordeaux build farm
>> nar storage or to take on the hosting of the data serivce instances, I'm
>> also how long the bordeaux build farm will keep going for. If it is
>> retired, the Guix Foundation already owns the Honeycombs so they can be
>> moved and repurposed for CI.
>
> I can feel the bitterness and understand it, but I have a different
> interpretation: it’s not that “there’s no concrete plan”, but rather
> that there’s apathy and inertia that constantly need to be fought.
> We’re the ones making plans concrete, after all; in this case, the FSF
> funds and Guix Foundation have the money, so no major obstacle other
> than inertia.

Which is why I'm saying if there's more need for ARM build machines for
CI, the HoneyComb machines currently in use for bordeaux could be used
for that instead. That's a quicker change to make than buying more
machines.

Why not reconfigure hamal at least and see how it performs?

signature.asc
Description: PGP signature