[Feature]: Add latency benchmarks #48

Activity

1. mawad-amd

   added

   benchmarks

   enhancementNew feature or requestNew feature or request

   on Jul 13, 2025
2. arcusbuilds commented on Oct 12, 2026

   

   arcusbuilds

   on Oct 12, 2026

   More actions

   Hi, I'd like to work on this. Before I start, some scope questions:

   1. Should this be a new example (e.g. examples/NN_latency/) or live under benchmark/?
   2. Ops: just load/store, or also get/put/atomics?
   3. Measure both local and remote (peer) latency, sweeping small message sizes (e.g. 4 B to 64 KB)?

   I'll add a test like tests/examples/test_jacobi.py too.
3. mawad-amd commented on Oct 12, 2026

   

   mawad-amd

   on Oct 12, 2026

   CollaboratorAuthor

   More actions

   Thanks @arcusbuilds!

   Benchmark rather than an example please — a new benchmark/latency/. No test needed, benchmarks only.

   Use iris.bench for the sweep: @bench.register with @bench.axis(...) gives you the grid, arg parsing and CSV for free. benchmark/ccl/bench_all_gather.py is a short example.

   I'd make everything an axis — message size, source rank, destination rank, and the op. Loads and stores are enough to start. Sizes 4 B to ~4 MB log-spaced.

   For the timing, use read_realtime() from iris/mem/utils.py (wraps s_memrealtime, 100 MHz constant clock) and read it either side of the op inside the kernel. iris/mem/triton/tracing.py shows the pattern.

   ib_write_lat is a good model worth copying — a ping-pong where half the round trip is the one-way latency, reporting min/median/99% rather than a mean. See run_iter_lat_write in src/perftest_resources.c.

   Let me know if you have questions.