I work on distributed runtimes and accelerator communication: collective algorithms, GPU runtime behavior, and the measurement code around them.
Mostly C++, CUDA, HPX/MPI, Linux systems work, and performance experiments that try to make the failure mode measurable.
HPX · C++ / HPX / MPI
56 merged upstream PRs. In hpx::collectives: hierarchical all_reduce,
all_gather, all_to_all, inclusive_scan, and exclusive_scan; unified
generation handling, flattened hierarchical payloads, hardened failure paths,
and size-based pairwise all_to_all.
The pairwise campaign covered 960 executions across 16–128 localities and 4 B–64 KiB per destination.
In the runtime: orderly MPI parcelport teardown, shutdown races in the TCP sender and background threads, Dijkstra termination-probe fixes, failure detection and killed-locality handling, the nvcc CUDA builds, and the LSU Rostam CI.
HAMi host-PID broker · Go / C / CUDA
The device plugin serves a Unix socket and reads each caller's host PID from
SO_PEERCRED, so HAMi-core finds it without the before/after CUDA
primary-context probe and without mounting host procfs. Server and device
plugin in HAMi#2417, init-time
client in HAMi-core#354.
Off by default; falls back to a node-wide lock and the old probe when the
broker can't answer.
HAMi-core · C / CUDA / Linux
Replaced a post-initialization lock that stayed stuck after its owner was
killed with process-recoverable POSIX record locking, plus concurrency
regression coverage. Smaller fixes for repeated RTLD_NEXT lookups returning
null, rate limiting on cuLaunchCooperativeKernel, and the fan index and
timestamps the v2 hooks were dropping.
CommCanary · Python
Source-verifiable distributed-communication canaries. Turns workload traces into replayable artifacts, then checks source correspondence, timing, rankings, and decision behavior instead of assuming a smaller trace preserved them.
Other upstream fixes
- triton#12001 · file cache groups keep working after the cache dir is moved
- dynamo#15395 · replayed tokens no longer counted as prompt tokens after a request migrates
GPU-aware communication, memory registration, runtime initialization, symmetric-memory collectives, transport progress, and how to make performance claims survive reproduction.