Live Q&A with Speakers – Joseph Provine
summary
Live Q&A reveals concrete systems and numbers: a deployment using 16×200G NICs injected into 8×400G uplinks to reduce ECMP hash collisions; GPUDirect peer memory via Mellanox peer direct for zero copy; RCQ pairs created before jobs for flow multiplexing; Rocket used for RCA and RDMA contraction analysis; Prime and Chakra traces feed a traffic generator; Arcadia simulator runs packet mode single-threaded (one day per simulated day at 12K GPUs) or flow mode (a year in hours), models PFC/ECN exactly or approximately, and is slated for open sourcing.
AI-written summary · about
chapters (20)
- 0:00Q&A session
- 0:25Faster fabric links and GPU traffic
- 1:50Dynamic topology enrichment
- 2:33Simulating AI workflow traffic
- 2:54GPU memory pinning and NIC
- 3:36Traffic patterns for AI workloads
- 4:17Detecting capacity imbalance
- 5:03Flow multiplexing with RCQ pairs
- 5:45Telemetry and observability trade-offs
- 6:52Arcadia open source direction
- 7:13Arcadia system requirements
- 7:56Deep buffer switches in AI
- 9:21Rocket common use cases
- 9:41Arcadia validation results
- 10:43ECMP hashing with QP IDs
- 11:26Ramcom replay for collective behavior
- 12:06Synthesized workload traces
- 13:12Network latency impact on AI
- 14:36Simulating PFC and ECN with Arcadia
- 15:00Closing remarks
