Mark Grondona

I wonder how much work it would be to properly support gpus in librlist? Perhaps that would then allow sched-simple to allocate them (currently it cannot). My other worry here...

Ah, probably dropped in `rlist_append` and that seems like a bug :-(

> but it doesn't look like sched-simple is returning queue properties in allocated Rs. Is it possible that was by design? I'm not sure, but I did think properties were...

Oof. I guess sched-simple is so slow at that scale it doesn't matter. I'd like to see results for something smaller as well just in case though.

The impact is not as bad as I'd feared! The rlist implementation is kind of dumb and not optimized at all, so we could look into improving it. This would...

> So my comment above was about shifting some of the work from the job manager to the resource module to reduce d the penalty on every allocation. That makes...

FWIW, I really like the approach.

I've been seeing that one regularly in CI, but haven't been able to reproduce locally. It appears the Python test script in the shell-lost test is getting sent SIGINT while...

> which doesn't make sense because there's a barrier in there 😕 D'oh, the libflux `flux_barrier()` returns a future, but the Python handle `barrier` method returns nothing. So this is...

Feel free to squash those!