On Fri, Sep 18, 2015 at 12:24:00PM -0400, Randy MacLeod wrote:
> On 2015-09-07 11:14 AM, Martin Jansa wrote:
> > Full jenkins build test takes over 3 days (depending on sstate reuse,
> > one MACHINE can take more than 24 hours), so if I add them to
> > master-next after it's merged to oe-core it will take 3-6 days to
> > build-test them (6 days if they are sent just after current queue is
> > sent for build).
>
> Hi Martin,

Hi Randy,

> Replying privately but you can add the list back in your
> reply if you like.

OK, adding both MLs

> I never build all of meta-oe but a day seems like a long time.
> On a 3-4 year old 24 core system, with 64 GB RAM and
> 2 large disks, I build oe-core + a whitelist of meta-oe + misc layers in
> 5 hours using 18.3 of the 24 cores on average.
> What are the specs of your build machine?

It's one of VMs running on box donated by HP to webos-ports project. It
has access to 8 E5-2630L 0 @ 2.00GHz cores (from IIRC 24) and slow IO
which we try to compensate a bit by using tmpfs for TMPDIR, VM has 80G
RAM, so I'm using 72G for tmpfs.

Tom: please fill-in the gaps, I don't know any details about the real HW
it's running on.

The same machine is also running some webos-ports builds sometimes.

BUT be aware that my world build is much bigger than what you're
probably building and changes in oe-core almost always ensure very low
sstate reuse.

e.g. last build with only small oe-core upgrade since previous build:
NOTE: Tasks Summary: Attempted 33283 tasks of which 5252 didn't need to be rerun and 33 failed.
NOTE: Build completion summary:
NOTE:   do_populate_sysroot: 21% sstate reuse (485 setscene, 1750 scratch)
NOTE:   do_package_qa: 12% sstate reuse (262 setscene, 1760 scratch)
NOTE:   do_package: 4% sstate reuse (76 setscene, 1750 scratch)
NOTE:   do_packagedata: 13% sstate reuse (281 setscene, 1750 scratch)
NOTE:   do_package_write_ipk: 12% sstate reuse (262 setscene, 1756 scratch)
NOTE:   do_populate_lic: 22% sstate reuse (521 setscene, 1806 scratch)

real    1485m40.070s
user    4166m49.125s
sys     2825m50.068s

BTW: you can see all this information in the world logs I'm sending to MLs
(but not all the builds make it to ML report).

> A decent build box these days has 32+ cores, 64+ GB RAM  and
> and SSD or RAID disk filesystem. The costs of such a system
> are significant but worth it given the number of people's time
> involved in meta-oe.

That's true, but I have to do with what I have available :).

> Also, it would be nice if OE had a few such machines each
> working on builds for a MACHINE to reduce your turn-around
> to more like a day or so. :)

Yes, it would be nice, I'm trying to get Tom King to set another VM like
this in another location (another HP donated server), hopefully he will
finish it soon.

It's all hosted together with OE infrastructure, bandwidth is limited, but
all I need for these builds is to fetch sources and upload the logs, there
are no package feeds or sstate from these builds usable outside, but now it
rsync the downloaded sources to sources.openembedded.org

If you know about some idling build servers, I would like to give them
something to do :).

Regards,

--
Martin 'JaMa' Jansa     jabber: Martin.Jansa@gmail.com