1315158 - Launch b-2008 and y-2008 instances on more than c4.2xlarge

Status: RESOLVED WONTFIX

(Release Engineering :: General, defect)

Description

[Nick Thomas [:nthomas] (UTC+12)]

In bug 1315061 were not able to start b-2008 and y-2008 instances on c4.2xlarge because spot prices were higher than our bid. A bigger bid worked around it partially, and prices came down. Ideally we'd just float over to a different instance type like we do for bld-linux64, eg https://github.com/mozilla-releng/build-cloud-tools/blob/master/configs/watch_pending.cfg#L124

Could consider this for tests too.

Comment 1

[Nick Thomas [:nthomas] (UTC+12)]

Could also look at us-west-1e. We're currently excluding that availability zone, but from the way the prices moved around I think c4.2xlarge should be available there too.

Comment 2

[Chris AtLee [:catlee]]

Q/grenade - are there any other instance types that would be suitable for use here? Would c4.xlarge work in a pinch?

Comment 3

[Rob Thijssen [:grenade (EET/UTC+0300)]]

it could work. i've never tried. aiu the diferences between c4.xlarge and c3.2xlarge are half the ram (8:16) and half the cpus (4:8). we don't use ephemeral drives on bb win builders so not having them makes no difference. if it doesn't work, we'll probably see timeouts due to having less memory and processing power.

Comment 4

[Amy Rich [:arr] [:arich]]

Not sure if c4.xlarge will work (it might lower performance enough that we actually time out on hg clones or something), but I think c4.4xlarge will probably work fine. Adding mark to the bug for commentary. Testing would prove it either way.

We could also start looking into the new AWS region in ... Ohio(?) and build up infra there or in usw1.

Comment 5

[Gregory Szorc [:gps]]

Mercurial clone performance shouldn't change much within the c4 instance types as long as the I/O settings are the same.

However, compiling Firefox will be proportional to the number of CPU cores available. I recommend going no lower than a c4.2xlarge, otherwise build times will take a really long time. We run c4.4xlarge in TC. Ideally we'd run c4.8xlarge. But we need to optimize non-compile parts of the build jobs before we do that, otherwise we waste a lot of cores.