Connectivity peakmem benchmarks - #1661
Conversation
ASV BenchmarkingBenchmark Comparison ResultsBenchmarks that have improved:
Benchmarks that have stayed the same:
|
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Drop the GridBenchmark template, whose full cached_grid setup was always overridden, and root MinimalGridBenchmark on CachedFixtures directly so the suite only touches the parent-preloaded topology fixtures. Pin numba threads around the tracemalloc samples, as the other suites do, since edge_node and face_edge are built by parallel kernels. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
dylannelson
left a comment
There was a problem hiding this comment.
Everything works as stated and seeing no concerns. Approving with a comment as it may be nice to import the connectivities from the conventions file as opposed to re defining them here. But if there's a good reason for re defining them feel free to resolve the comment and push forward. We talked about potentially changing some of the conventions file so this would allow us to change in only one place rather than multiple, but that's a bit further down the line
Sevans711
left a comment
There was a problem hiding this comment.
Read through the changes and they all look reasonable. It's a bit strange to pass something around by editing a class attribute instead of an instance attribute, but the docstrings/comments clearly explain the motivation, that this decision was made to help deal with one of ASV's many quirks, so I'm happy to let this proceed as-is. It also looks like the connectivity benchmark results are all working properly, based on the ASV bot comment in this PR. Approving!
Closes #1660
Overview
Two sets of memory benchmarks, tracking allocations in isolation and total peakmem usage.
ConnectivityTracemalloc results from Derecho:
(face_edge is a flat 11.4k, face_node is a flat 2.54k)
PR Checklist
General
Testing & Benchmarking
Documentation
AI Disclosure
AI Usage: Claude Opus 5