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
- 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
- RotaSched:OS 式 proactive rotary scheduling,用 Virtual Lag Time 按 TTFT/TBT 紧迫度轮转请求
- DuplexKV:合并碎片化 paged KV、全双工无竞态传输、与模型执行 overlap
- SLO-aware memory co-design:突破仅 priority reorder 的 SLO 调度,在 superchip 上缓解 HOL blocking
- 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 / throughput | Comparable to baselines |
与 wiki 交叉引用
- Inference Capacity Trap — KV 压力下 batch/排队与 SLO 冲突
- vLLM — paged KV 基线;SuperInfer 解决 scatter 迁移开销
- Heterogeneous Inference — GPU–CPU 异构 offload 谱系
- Prefill Decode Divergence — TTFT vs TBT 双 SLO 维度
- Disaggregated Inference — 服务层调度与内存分层
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