[FEA] Add support for RMM logging to rapids-pytest-benchmark

[rlratzel]

The new API is here:
rapidsai/rmm#363

• The rapids-pytest-benchmark code can effectively replace the code in gpu_metric_poller.py with a similar module responsible for reading the log file.
• Since RMM logging adds significant overhead, the benchmark time measurements cannot be run when logging is enabled. This is the same case as GPU polling, so the approach already used in rapids-pytest-benchmark should not change (ie. separate runs done internally).
• The peak memory usage is what we want to see, since this shows what the resource requirements of our algos are, and we obviously want them to be as low as possible to allow customers to run large datasets.
• (need to confirm this detail) The log contains lines/entries for each alloc and how many bytes were requested, and each free and how many bytes were freed. Algos make several alloc/free calls throughout their lifetime, so the log could get large.
• (this is based on the assumption above, which needs to be confirmed) To compute the peak memory usage, do the following:
  • initialize vars max and total to 0.
  • for each alloc, add the number of bytes to total. If total > max, set max = total
  • for each free, subtract the number of bytes from total
  • at the end of the benchmark run, return max as peak memory usage
  • bonus: at the end of the benchmark run, total should be 0. If not, report total as the number of leaked bytes.

[rlratzel]

@dillon-cullinan is working on this and is planning on having it done post-MVP timeframe.

[ajschmidt8]

adding a link to the relevant rmm issue below for reference

rapidsai/rmm#423

[ajschmidt8]

Blocked by rapidsai/rmm#721

[ajschmidt8]

Unblocked with the merge of rapidsai/rmm#722

[ajschmidt8]

#62 is merged which closes this issue.

0.0.14 packages have been built and uploaded to Anaconda.org:

• https://anaconda.org/rapidsai/rapids-pytest-benchmark
• https://anaconda.org/rapidsai-nightly/rapids-pytest-benchmark