Reproduce “The Semantic Arrow of Time, Part III: RDMA and the Completion Fallacy” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Paul Borrill
Paul Borrill's paper rigorously identifies the "completion fallacy" in Remote Direct Memory Access (RDMA), demonstrating that its completion signals do not ensure application-level semantic agreement, leading to systematic semantic corruption in high-performance distributed systems. The analysis, supported by production-scale case studies from major tech companies, argues this fundamental design flaw undermines reliability, particularly in AI training workloads.
Key insights
- The "completion fallacy" establishes that RDMA's completion signal (T4) does not guarantee data visibility to the remote CPU (T5) or application-level semantic agreement (T6), leaving a critical and potentially large temporal gap.
- RDMA's communication model is fundamentally "forward-in, then-out" (fito), lacking the necessary "reflecting phase" required by the oae framework for explicit semantic commitment.
- The atomicity gap, where RDMA's atomic operations are limited to 8 bytes, is insufficient for atomically updating complex, multi-field data structures, leading to internal semantic corruption.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.