← All publications
ISCA2025

A4: Microarchitecture-Aware LLC Management for Datacenter Servers with Emerging I/O Devices

Haneul Park, Jiaqi Lou, Sangjin Lee, Yifan Yuan, Kyoung Soo Park, Yongseok Son, Ipoom Jeong, Nam Sung Kim

A4 manager CPU cores Shared LLC CPU ways I/O ways non-inclusive cache directory hidden I/O ↔ CPU contention 100+ GbE NIC DCA (DDIO) NVMe · CXL dev high-speed I/O DCA writes

Emerging I/O devices inject data straight into the shared LLC via direct cache access, colliding with CPU workloads — and with each other — inside structures software cannot see.

TL;DR

High-speed I/O devices and CPU cores contend for the last-level cache in ways rooted in the microarchitecture — not the ISA. A4 identifies these hidden interference channels and manages LLC allocation around them for datacenter servers.

Why it matters

In modern server CPUs, the last-level cache serves not only as a victim cache for the private caches but also as a buffer for low-latency DMA transfers between cores and I/O devices through direct cache access (DCA). Prior work showed that high-bandwidth network I/O can flood the LLC with packets and contend with co-running workloads — but the story runs deeper than that.

Key findings

BibTeX

@inproceedings{park2025a4,
  author    = {Haneul Park and Jiaqi Lou and Sangjin Lee and Yifan Yuan
               and Kyoung Soo Park and Yongseok Son and Ipoom Jeong and Nam Sung Kim},
  title     = {{A4}: Microarchitecture-Aware {LLC} Management for Datacenter
               Servers with Emerging {I/O} Devices},
  booktitle = {Proceedings of the 52nd Annual International Symposium on
               Computer Architecture (ISCA)},
  year      = {2025}
}