Run MakeBinaryPackage on r6i.4xlarge EC2 instances instead of m5.4xlarge

[Atry]

According to https://aws.amazon.com/ec2/instance-types/, r6i.4xlarge doubles the memory in comparison to m5.4xlarge. Hopefully it would reduce the out of memory errors.

[Atry]

I have deployed the step functions for test

[Atry]

I triggered a previously failed build bin/build-on-aws 2022.01.21 debian-11-bullseye for testing this PR.

[Atry]

Sorry, deploying the step functions do not take effect.

I just deployed the lambdas instead, and triggered a previously failed build bin/build-on-aws 2022.01.20 debian-11-bullseye for testing this PR.

[Atry]

I triggered a previously failed build bin/build-on-aws 2022.01.21 debian-11-bullseye for testing this PR.

The retry instance type is r6i.4xlarge since I have deployed new lambdas.

[fredemmott]

Nice, the m* used to be the high memory tier, and I assumed that had continued

Given we don't need these jobs to be super fast, we should also try lower options in the r6i range - changing the cores:RAM usage should fix the problem - it's not just a matter of more RAM.

Worth trying r6i.xlarge, and seeing what the build times are like.

--

As a concrete example, we were building with 32GB up until August last year (a4c60f8) until some OOMs started - doubling the cores and doubling the RAM did not fix the problem.

[Atry]

There is one retry attempt for the nightly build. The number of retry attempts is less than previous nightly builds.

[Atry]

Currently we limited the concurrency in #260.

Let's revert #260 for better CPU utilization

[fredemmott]

There is one retry attempt for the nightly build. The number of retry attempts is less than previous nightly builds.

If it still needs an OOM retry, the problem exists; sometimes, with the previous settings, it succeeds with 0 retries. "Fewer retries" is extremely low signal.

I think the next step is to enable atop to figure out what's going on; I suspect cargo may be parrelising to NCORES in combination with HHVM's own parellization, but we need more data

[Atry]

Even though it's a weak signal, we did find less failures recently. Shall we merge this PR?