SuperInfer: SLO-Aware Rotary Scheduling on Superchips

Authors: Jiahuan Yu, Mingtao Hu, Zichao Lin, Minjia Zhang | Affiliation: UIUC | PDF: SuperInfer_SLO_Aware_Rotary_Scheduling_Superchips_2026.pdf

一句话总结

SuperInfer 为 GH200 等 superchip(NVLink-C2C ~900 GB/s)联合设计 RotaSched(按 SLO 进度主动轮转 KV 于 HBM/DRAM)与 DuplexKV(全双工 KV 迁移引擎),高负载下 TTFT SLO 达成率最高 +74.7%,且 TBT/吞吐与 SOTA 相当。

核心贡献

  1. Superchip 瓶颈诊断:直接移植 PCIe offload 仅利用 <5% C2C 带宽 — 根因在软件栈而非硬件
  2. RotaSched:OS 式 proactive rotary scheduling,用 Virtual Lag Time 按 TTFT/TBT 紧迫度轮转请求
  3. DuplexKV:合并碎片化 paged KV、全双工无竞态传输、与模型执行 overlap
  4. SLO-aware memory co-design:突破仅 priority reorder 的 SLO 调度,在 superchip 上缓解 HOL blocking
  5. GH200 全栈评估:高负载 TTFT SLO +74.7%;低负载无退化

关键数字

设置结果
TTFT SLO attainment (high load)Up to +74.7% vs SOTA
NVLink-C2C bandwidth~900 GB/s (vs PCIe 32–64 GB/s)
PCIe offload on GH200<5% C2C utilization
TBT / throughputComparable to baselines

与 wiki 交叉引用

Citations

[1] SuperInfer_SLO_Aware_Rotary_Scheduling_Superchips_2026.pdf — Yu et al. (2026) [2] superinfer-slo-aware-rotary-scheduling.md — 结构化摘录