Workshop: 6th International Workshop on RESource DISaggregation in High Performance Computing (RESDIS)
Authors: Amirreza Barati Sedeh, Ryan Eric Grant, and Ahmad Afsahi (Queen's University)
Abstract: Traditional server-centric data center design relies on a fixed ratio of compute and memory, limiting adaptability to evolving workloads, particularly in AI. Composable Disaggregated Infrastructure (CDI) addresses this by enabling on-demand composition of disaggregated devices. We focus on the intersection of GPU disaggregation and MPI collective communications, introducing an emulation methodology that uses available hardware to emulate PCIe-based disaggregated GPUs. While high-end servers run collectives over high-bandwidth intra-node interconnects, PCIe-based disaggregated systems introduce congestion, making it challenging to identify the best-performing algorithm for each collective. We propose a congestion-aware scoring model that estimates congestion at step granularity and, grounded in a Hockney-based bandwidth model, predicts the absolute communication latency of each algorithm on different PCIe topologies. Experimental results show the model predicts absolute latency within an average accuracy of 94.73%, 93.25%, and 93.94% for Allgather, ReduceScatter, and Alltoall, respectively, providing an effective, profiling-free tool for algorithm selection in GPU-disaggregated systems.
Back to 6th International Workshop on RESource DISaggregation in High Performance Computing (RESDIS) Archive Listing
Back to Full Workshop Archive Listing