Performance Degradation on Independent Task benchmark

[wlruys]

I've been trying to reproduce our scaling plots from February to show the advantages of the C++ runtime for TST.

I've found our scaling is highly degraded for 1000 1ms independent tasks.
On a RTX node, we only have 4.5x speedup on 8 threads.
On the SKX node, we get 5.9x speedup (down from 7.4x speedup in Feb.)

Disabling contexts gets us back to 6.57x speedup.
Using a single C++ call for cleanup instead of 2 (PARLA_ENABLE_PYTHON_RUNAHEAD=false), gets us to 6.9x speedup.

I'm not yet sure where the remaining missing time is. If its on the python side, possibly in creating device requirements?

[wlruys]

We need to make our benchmarks easy to reproduce and automated again. (e.g. fix the google benchmark scripts) and to check them before any feature merges.

[nicelhc13]

questions:

1. is it a single-device task or multi-device task?
2. is it using any data operand?
3. it should be cpu right? (if u tested on skx node)

i agree with the google benchmark. the issue of the google bench is that it breaks binlog.

[wlruys]

1. single
2. no data
   3 yep this was cpu only throughput testing

I'll handle this since its really most of my code that hurt it. I need to move parts of contexts to c++ and decrease the amount of new dictionary allocations on the python side.