remove superfluous ABT_thread_yield() call

[carns]

improves ops/s by 2-3% in sequntial benchmarks

[mdorier]

Careful here; this ABT_thread_yield was added to solve this: #278

[carns]

Ah, shoot. I keep forgetting about the desire to support the use case where execution streams are mapped to multiple pools.

I'll take a closer look. If that's the case (that we need to do a blind yield, even if there is nothing to do in the current pool, in case there is another pool to be serviced) then its the other yield call at https://github.com/mochi-hpc/mochi-margo/blob/main/src/margo-core.c#L1984 that needs to be removed.

It doesn't make sense to do a yield and then a few lines further down in the function yield again, so one of them needs to go. It sounds like we need to cut the one that has guards around it checking the current pool and go with the one that yields no matter what.

I haven't checked the Argobots code, but for some reason these yields have a measurable cost in low latency operations even if there is nothing to yield to. I must be the cost of pushing the work unit back into the pool (possibly checking other pools depending on the ES configuration) and then popping it back out.

[carns]

Trying another strategy with this commit (and also more carefully commenting the reasoning for the logic). If this passes CI then I think it will work.

[codecov]

Codecov Report

All modified and coverable lines are covered by tests ✅

Project coverage is 57.86%. Comparing base (5f08e35) to head (9854e4e).

Additional details and impacted files

@@            Coverage Diff             @@
##             main     #293      +/-   ##
==========================================
+ Coverage   57.80%   57.86%   +0.05%     
==========================================
  Files          70       70              
  Lines       10168    10165       -3     
  Branches     1335     1334       -1     
==========================================
+ Hits         5878     5882       +4     
+ Misses       3456     3451       -5     
+ Partials      834      832       -2     

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

[carns]

Yeah this variation seems to work fine. Squashing and merging. Thanks for pointing out the problem with the original attempt @mdorier