Event Actions
Characterizing and Mitigating Communication Overheads in Large-Scale Accelerator Systems
Abstract:
Traditional approaches to improving processor performance have become unsustainable, leading to the integration of domain-specific accelerators to achieve substantial gains in performance and energy efficiency compared to general-purpose processors. However, these accelerators are now reaching their limits in addressing the growing computational and memory demands of emerging applications. The natural progression is to scale these accelerators to meet these demands. While scalability offers significant computational and memory advantages, it is constrained by communication bottlenecks caused by slow interconnects between devices.
Our research addresses this critical challenge through two key strategies. First, we propose optimizations that shift communication overhead off the critical execution path, reducing its impact on performance. Second, we enhance the efficiency of network interconnects by optimizing bandwidth utilization, thereby narrowing the gap between intra-device accesses and inter-device communication.
Committee:
- Kevin Skadron, Committee Chair (CS/SEAS/UVA)
- Adwait Jog, Advisor (CS/SEAS/UVA)
- Brad Campbell (CS,ECE/SEAS/UVA)
- Mircea Stan (ECE/SEAS/UVA)