Seunghyun An, Joontaek Oh, Ming Liu
- Found that Linux’s conventional scheduling signals—CPU load and static topology/locality—break down on chiplet CPUs: under shared I/O-chiplet congestion, topologically local placement can run 4.3–4.5× slower than remote placement
- Isolated this “I/O Chiplet Wall” from memory-controller saturation, showing that the slowdown originates from contention along the compute-to-I/O-chiplet path itself
- Built cSwitch, which uses commodity performance counters to infer each thread’s communication path and bandwidth demand, then schedules based on runtime path congestion rather than static topology
Seunghyun An, Joontaek Oh, Ming Liu
- Characterized the chiplet network of two generations of AMD EPYC CPUs
- Proposed future directions with chiplet networking
