SuperPMI: Improve superpmi-diffs pipeline parallelism: split asmdiffs and tpdiff

[BruceForstall]

Both asmdiffs and tpdiff are done for a single platform consecutively on the same Helix machine. In a recent run, it seems each takes about 30 minutes (MCH download takes 6 to 15 minutes).

If tpdiff ran on its own Helix machine, it could run in parallel, improving overall pipeline throughput. Also, asmdiffs only needs to depend on the Checked JIT builds, and tpdiff on the Release JIT builds.

Doing this would probably require altering how the diffs "summarizing" happens with superpmi_diffs_summarize.py.

category:eng-sys
theme:super-pmi
skill-level:intermediate
cost:medium
impact:medium

Tagging subscribers to this area: @JulieLeeMSFT, @jakobbotsch, @kunalspathak
See info in area-owners.md if you want to be subscribed.

Issue Details

---

Both asmdiffs and tpdiff are done for a single platform consecutively on the same Helix machine. In a recent run, it seems each takes about 30 minutes (MCH download takes 6 to 15 minutes).

If tpdiff ran on its own Helix machine, it could run in parallel, improving overall pipeline throughput. Also, asmdiffs only needs to depend on the Checked JIT builds, and tpdiff on the Release JIT builds.

Doing this would probably require altering how the diffs "summarizing" happens with superpmi_diffs_summarize.py.

category:eng-sys
theme:super-pmi
skill-level:intermediate
cost:medium
impact:medium

Author:	BruceForstall
Assignees:	-
Labels:	area-CodeGen-coreclr
Milestone:	Future

[BruceForstall]

Extending this idea: we could create even more parallelism, such as a single Helix job for each separate <MCH file, tpdiff/asmdiffs type> pair. That is, for each MCH file, create one Helix job for tpdiff, and one for asmdiffs. This creates a separate job for each superpmi.py invocation. We would need to ensure the correlation payload (copied to every Helix machine) is minimized to minimize overall overhead.

[BruceForstall]

This is basically complete -- the ideas above can be revisited if we decided throughput needs to be improved more.