CCCL 3.2 (and 3.1) benchmark on B200

[bernhardmgruber]

I figured we could use an up-to-date benchmark of CUB and Thrust, so I ran a few batch jobs on a B200. The commit c5a7406 is just before we branched 3.2, so if there are no perf related backports, this will correspond to CCCL 3.2:

CUB:

Thrust:

Benchmark SQLite database (on CCCL team drive):
https://drive.google.com/drive/folders/1eaRnrsz7gEmX386rgI0L_boRJIejNGDk

The benchmark was run using benchmarks/scripts/submit_benchmark_job.sh, which runs run.py. The plot is made with sol.py (ran twice with a regex to split CUB and Thrust).

CUB 3.1 (blue) vs CUB 3.2 (orange):

[bernhardmgruber]

A few things I noticed that could be improved:

• For CCCL 3.0, I remember the total benchmark time was sometime around 2-3h on a B200. Now it was >4h so I needed two batch jobs. I think some benchmarks take extraordinary long (e.g. the segmented ones), and we should consider reducing the total number of variants to test. Alternatively, I could have also run the P0 subset. Issue: Some benchmarks take really long #6731
• Some benchmark names (the application inspired ones) have some really long names
• Some benchmarks do not report a bandwidth, which crashes the sol.py script and I had to manually drop them from the SQLite database. See also: sol.py fails to plot some benchmarks #6729