Jumanji - NUCA-Aware Data Placement for Latency-Critical Applications

Tuesday April 9, 2019
Location: CIC Panther Hollow
Time: 4:30PM-5:30PM

Prior work has shown that data placement in the shared LLC strongly impacts processor performance on non-uniform cache access (NUCA) architectures. However, previous data placement techniques focus only on minimizing total data movement. This optimization metric actually harms other application objectives, specifically tail latency of user-facing applications. Additionally, previous techniques for meeting tail latency deadlines are inefficient because they lack benefits from NUCA-aware data placement. We propose Jumanji, the first data placement technique which optimizes for the colocation of latency-critical and batch applications. Jumanji’s simple hardware and intuitive data placement algorithm minimize data movement for batch applications without violating tail latency deadlines.

Brian Schwedock is a second year ECE PhD student advised by Professor Nathan Beckmann. His research encompasses computer architecture and computer systems with a strong focus on caching. His projects target caching applications for data center workloads.