| name | smetric-llm-scheduling-serving-agents-session-centric |
| description | Balanced session-centric LLM scheduling for agent serving workloads. Routes first request in each session for load balance and follow-ups cache-aware. 10-16% TPS improvement. Leverages intra-session locality and 80%+ KV-reuse in agent traces. Activation: LLM scheduling, agent serving, session-centric scheduling, inference infrastructure, tokens-per-second. |
| metadata | {"arxiv_id":"2607.08565","published":"2026-07-09","authors":"Jiahao Wang, Kaizhan Lin, Kaixi Zhang, Jinbo Han, Xingda Wei","tags":["llm-scheduling","agent-serving","session-centric-scheduling","inference-infrastructure","tokens-per-second"]} |
SMetric: Rethink LLM Scheduling for Serving Agents with Balanced Session-centric Scheduling
Overview
LLM scheduling is critical to serving, yet existing designs are optimized for human-issued requests, not agent-issued ones. Agent serving shifts the workload: agents act on complete responses (making TPS the primary goal) and requests share much of their KV-cache (80%+ reuse in production traces). SMetric addresses the tension between KV-reuse and load balance with balanced session-centric scheduling.
Key Innovations
Agent Serving Characteristics
- Agents act on complete responses — TPS is primary, per-token latency is relaxed
- KV-reuse exceeds 80% of request tokens in agent traces (vs 54-62% in chat)
- Existing schedulers over-prioritize routing to cached instances, capping TPS
Balanced Session-Centric Scheduling
- Routes each session's first request purely for load balance
- Follow-up requests routed cache-aware (preserving local KV-reuse)
- Session turn information derived from user inputs alone — scheduler stays clean and stateless
- Balances a small fraction of requests (first in each session) to balance the cluster
Two Key Insights
- Load balance need not sacrifice all KV-reuse (thanks to global-tier KV store)
- Intra-session locality means balancing first requests suffices for cluster balance
Methodology
- Session Detection: Identify session boundaries from request patterns
- First-Request Routing: Route first request in session for load balance
- Follow-Up Routing: Route subsequent requests cache-aware
- Stateless Design: Use session turn info from user inputs, no global state
Implications
- Agent serving requires fundamentally different scheduling than human serving
- Session-centric scheduling is a simple yet powerful abstraction
- Stateless design avoids the complexity of global session tracking
- 10-16% TPS improvement demonstrates real-world impact
Pitfalls
- Session detection from user inputs may not be perfectly accurate
- Global-tier KV store availability affects performance gains
- Workload characteristics may vary across different agent applications
- First-request routing may increase cold-start latency for new sessions
Activation Keywords
LLM scheduling, agent serving, session-centric scheduling, inference infrastructure, TPS optimization, KV-cache reuse, load balancing, SMetric, BAILIAN trace
Paper Reference
arXiv:2607.08565 - "SMetric: Rethink LLM Scheduling for Serving Agents with Balanced Session-centric Scheduling" (Jul 2026)