fix: Add benchmarks for loaded params

[winston-h-zhang]

This PR adjusts the fibonacci bench to also test loaded params. In the benchmark, we first test with a set of freshly generated params, writing them to disk if necessary, and then we test again with a loaded set of params.

[winston-h-zhang]

!gpu-benchmark

[github-actions]

Benchmark for 3b9c434

Click to view benchmark

Test	Base	PR	%
LEM Fibonacci Prove - rc = 100/fib/num-100	6.0±0.06s	N/A	N/A
LEM Fibonacci Prove - rc = 100/fib/num-200	12.3±0.08s	N/A	N/A
LEM Fibonacci Prove Fresh - rc = 100/fib/num-100	5.9±0.07s	N/A	N/A
LEM Fibonacci Prove Fresh - rc = 100/fib/num-200	12.7±0.03s	N/A	N/A
LEM Fibonacci Prove Load - rc = 100/fib/num-100	6.0±0.07s	N/A	N/A
LEM Fibonacci Prove Load - rc = 100/fib/num-200	12.6±0.02s	N/A	N/A

[winston-h-zhang]

!gpu-benchmark

[github-actions]

Benchmark for 3b9c434

Click to view benchmark

Test	Base	PR	%
LEM Fibonacci Prove - rc = 100/fib/num-100	3.9±0.06s	N/A	N/A
LEM Fibonacci Prove - rc = 100/fib/num-200	8.0±0.27s	N/A	N/A
LEM Fibonacci Prove Fresh - rc = 100/fib/num-100	3.9±0.03s	N/A	N/A
LEM Fibonacci Prove Fresh - rc = 100/fib/num-200	8.5±0.14s	N/A	N/A
LEM Fibonacci Prove Fresh - rc = 600/fib/num-100	3.1±0.01s	N/A	N/A
LEM Fibonacci Prove Fresh - rc = 600/fib/num-200	7.0±0.01s	N/A	N/A
LEM Fibonacci Prove Load - rc = 100/fib/num-100	3.9±0.02s	N/A	N/A
LEM Fibonacci Prove Load - rc = 100/fib/num-200	7.8±0.06s	N/A	N/A
LEM Fibonacci Prove Load - rc = 600/fib/num-100	3.1±0.01s	N/A	N/A
LEM Fibonacci Prove Load - rc = 600/fib/num-200	7.0±0.03s	N/A	N/A

[winston-h-zhang]

The strangest thing is that there seems to be no slowdown on this GPU with loaded params.

[arthurpaulino]

I'm concerned about the increase in cost for this benchmark. Can we measure it in another place?

[huitseeker]

All of our benchmark machines are configured to not let the application write anything to disk, and do so without letting it fail to read or write. The bench machines send the bytes to /dev/null and successfully read zero-sized files.
(for more details : https://github.com/abbbi/nullfsvfs )

As a consequence, modifying the fibonacci benchmark is not a good way to test for public parameter loading.

[winston-h-zhang]

Ah, thank you for the heads up. That explains the absence of any slowdown, then. I guess I'll have to manually test any loading-related regressions.

[huitseeker]

Leaving this in request changes for clarity.

[winston-h-zhang]

Closed after argumentcomputer/arecibo#137