Conspirator - SmartNIC-Aided Control Plane for Distributed ML Workloads

Explore a groundbreaking conference talk from USENIX ATC '24 that introduces Conspirator, an innovative control plane design for distributed machine learning workloads. Delve into how this novel approach leverages SmartNICs to address CPU bottlenecks and suboptimal accelerator scheduling simultaneously. Learn about Conspirator's ability to facilitate efficient data transfer without host CPU involvement and its integration of a new scheduling algorithm that adapts to heterogeneous accelerators and changing workload dynamics. Discover the significant improvements Conspirator offers, including a 15% reduction in end-to-end completion time compared to RDMA-based alternatives, 17% better cost-effectiveness, 44% improved power efficiency, and a 33% reduction in GPU hours through optimized scheduling decisions. Gain insights into the evolving role of SmartNICs and their potential to revolutionize distributed ML workload management in this 18-minute presentation by researchers from Northwestern University and Hewlett Packard Labs.