This paper targets Molecular Dynamics (MD) workloads running on Wafer-Scale Chips (WSCs). It exploits key properties of both: MD communication is spatially local, phased, and highly synchronous; WSC-based MD systems use stride-based PE injection scheduling (“wavelet scheduling”) to limit concurrent traffic injection. Together, these properties create a unique opportunity that’s unavailable for arbitary NoC traffic: a PE’s local router credits become a reliable proxy for queue buildup, caused by its traffic, along the first few downstream hops. The reason is that downstream congestion backpropagates through CBFC, while a sufficiently large stride suppresses other-PEs generated traffic that’d make the credit signal noisy; MD’s spatial locality keeps the communication paths, and thus relevant congestion, within nearby downstream hops. WaveCC then uses these local credit levels and gradients to adapt the global stride and local injection rate, respectively, reducing backpressure-induced stalls and tail communication latency.

Comments

  • Intersting paper presenting workload–network co-design, where under MD’s fixed/local communication pattern plus stride-based injection scheduling, a normally very local and noisy signal (local router’s credit counts) becomes informative about multi-hop congestion.