On Mon, Aug 8, 2016 at 2:38 PM, Gregory Szorc <g...@mozilla.com> wrote:
> Bug 1290282 landed on the autoland repo. It transitioned all our Linux
> TaskCluster tasks from c3/m3/r3.2xlarge instances to c4/m4.4xlarge
> instances. This means:
>
> * Builds now have access to 16 vCPUs instead of 8
> * Builds are running on Haswell instead of Ivy Bridge and are cycle for
> cycle faster
> * We're using EBS instead of local instance storage
>
> In addition,the TaskCluster worker pool has been consolidated to just
> "gecko-{{level}}-b-linux." Before, we had {dbg,opt}-linux{32,64} worker
> types. In the short term, this will likely lead to a lower TC worker cache
> hit rate and slower builds than if we had a cache hit. i.e. we'll do
> clobber builds more frequently. However, I don't believe this is enough to
> offset the gains from adding more CPU. And, once sccache2 lands, the
> importance of objdir cache hits should wane. Also once sccache2 lands, we
> may want to reevaluate our instance choice, as we may not need as many
> cores any more.
>
> There's also bug 1291940 which is sapping some performance from the
> additional cores. That's the primary reason we can't use c4.8xlarge
> instances at this time.
>
> I don't anticipate this change will break anything. But we all know how
> sensitive our automation can be. Hence this email.
>
As a quick follow-up: I messed up the TaskCluster cache path configuration
and inadvertently caused different build configurations to share the same
cache. This has been causing intermittent build failures in automation. Bug
1293717 should hopefully land on Autoland soon and clear up this mess.
_______________________________________________
dev-builds mailing list
dev-builds@lists.mozilla.org
https://lists.mozilla.org/listinfo/dev-builds
