SMetric: Rethink LLM Scheduling for Serving Agents with Balanced Session-centric Scheduling

agent 2607.08565 — Cross-paper Synthesis

SMetric (2607.08565) — L3 per-paper synthesis #

Target: SMetric — balanced session-centric scheduling for agentic LLM serving. Peers in scope: RouteLLM (2406.18665), Autellix (2502.13965), Agent-protocol survey (2505.02279), Halo (2509.02121), KVCOMM (2510.12872), TokenCake (2510.18586), CPU-centric characterization (2511.00739), Continuum (2511.02230).

All nine entities sit in category: agent, but they split into a serving-systems cluster (the true peers of SMetric)

and a non-serving fringe that is related only by domain framing.

Tight cluster — cluster-level LLM serving for agents (share SMetric's exact problem space):

Related-by-diagnosis (empirical peers, not method peers):

Fringe (domain-adjacent, orthogonal method):


2. 本篇 vs 相关论文的 delta — new / incremental / contradictory #

What is genuinely new in SMetric (vs the cluster):

  1. Stateless session-turn inference as a scheduling signal. This is the single hardest-to-replicate contribution
  2. [2607.08565]. Every peer that exploits session/program structure keeps state: Autellix maintains a

    global process table [2502.13965], Continuum maintains a Tool Call Handler with per-tool

    history $S[f]$ [2511.02230], TokenCake requires a frontend graph registration API where users

    declare the agent DAG [2510.18586], and Halo requires the full workflow DAG up front

    [2509.02121]. SMetric alone infers turn purely from the count of historical messages

    already carried in the OpenAI-style request, needing no session→instance table, no client-side API change, and no

    garbage collection as sessions silently idle. This is a real paradigm difference, not an increment.

    1. "Balance only the first request" as a sufficient lever. The empirical foundation — 96.6% of follow-ups return to
    2. the first-turn instance under cache-aware routing vs only 4.0% under load-balancing

      [2607.08565] — is a mechanism no peer articulates. Autellix's locality-aware routing has the same

      effect (session stickiness) but treats it as a load-balancer heuristic, not as the thing that reduces the balancing

      problem to a small fraction of requests.

      1. Decoupling reuse from placement via the global tier. SMetric's Insight 1 (a KV-agnostic load-balancer comes
      2. within 7% of the best baseline once the global tier is fully provisioned) [2607.08565] reframes the

        whole reuse-vs-balance trade-off. TokenCake and Continuum both treat KV as pinned to one instance (offload to CPU

        and back, or TTL-pin in place); neither exploits a cluster-wide global tier that lets a re-routed session recover

        its cache remotely.

        Incremental / shared with peers:

        • The workload-study-then-scheduler structure (SMetric §3 Findings → §4 design) mirrors 2511.00739's
        • characterize→optimize pipeline [2511.00739]; SMetric's contribution

          over it is the specific findings (>80% reuse, 67% intra-session, ~90% reuse within 100s).

        • The "agent acts on complete response, so relax per-token latency, maximize TPS" objective shift
        • [2607.08565] echoes Autellix's move from per-call to program-level throughput

          [2502.13965] and Continuum's per-job (not per-token) delay metric [2511.02230]. SMetric

          states it most crisply but is not first to the idea.

        Contradictory:

        • SMetric's entire premise is that the GPU-side / KV-reuse imbalance caps TPS. 2511.00739 argues the opposite for a
        • different workload mix: that CPU tool execution is the E2E bottleneck (up to 88%), so GPU-side scheduling gains are

          capped at ~12% [2511.00739]. See §3 for the reconciliation.


        3. 可攻击面 — adversarial rebuttal against specific claims #

        Attack 1 — "the global tier decouples reuse from placement" over-generalizes from a favorable testbed.

        SMetric's Insight 1 rests on the global tier serving fetches "in time," and the paper itself admits the fully

        provisioned tier is "never the bottleneck" only for the model analyzed in §4.1

        [2607.08565]. The 235B experiment already shows the load-balance-only baseline losing 32% TPS because

        "a larger model puts more pressure on the global tier" [2607.08565]. On 200 Gbps RDMA NICs and 4 servers, the

        global tier is generous; on a bandwidth-starved cluster (or one without GPU-direct RDMA), the 7% "reuse ≈ placement"

        result could invert. TokenCake's own numbers reinforce the risk: even an in-node offload/upload round-trip is ~63.7 ms

        for 4096 tokens [2510.18586] — a remote fetch under contention would be worse, directly eating the

        "small TTFT bump agents tolerate."

        Attack 2 — the stateless turn signal is fragile to real agent behavior.

        SMetric infers turn from the number of historical messages [2607.08565]. But KVCOMM shows that in

        multi-agent workflows agents rewrite and compress context (compaction requests appear even in SMetric's own Figure 7),

        and forked sub-agents share a parent's context [2510.12872]. When the carried history is compacted or

        forked, "message count" no longer maps cleanly to "turn," so est_hit can mis-fire and force needless rebalancing. The

        paper's benign-failure argument ("dropped history → treat as first request, which has little reuse anyway") does not

        cover the compaction case, where the request genuinely does have large reuse but a shortened message list.

        Attack 3 — the load-balance metric is consistently the worst.

        In every reported setting SMetric's max/mean load ratio is worse than pure load-balance-only (2.6× vs 2.0×

        colocation-30B; 1.6× vs 1.3× 235B) [2607.08565]. SMetric's defense is that TPS, not balance, is the goal

        — but Autellix's Fig. 18 shows a visible gap to the SRPT optimal even for its throughput metric

        [2502.13965], and SMetric never bounds how far its greedy turn-gated policy sits from an offline

        optimal dispatch. The 235B "close ratios are partly a setup artifact (only 8 instances)" admission

        [2607.08565] suggests the balance advantage may shrink further at the small-cluster scales many providers actually run.

        Attack 4 — "no formal guarantee" is a shared but real weakness.

        Like Autellix (no competitive ratio vs SRPT) [2502.13965] and Continuum (no convergence proof for

        $\tau^*$) [2511.02230], SMetric offers no bound on global-tier fetch demand as a function of the

        first-request fraction — the very quantity its scalability argument depends on

        [2607.08565]. The claim "first requests are few, so the global tier suffices" is asserted from trace

        statistics, not derived; a workload with many short single-turn sessions (higher first-request fraction) is an

        un-analyzed adversarial regime.


        4. 生态位 — paradigm-shift positioning, adoption evidence #

        SMetric occupies the stateless-router-for-agentic-serving niche, and its positioning is best understood on a

        2×2 of (state kept: none ↔ full) × (control axis: spatial routing ↔ temporal retention):

        • Spatial + statelessSMetric (this niche is essentially unoccupied among peers; that is its moat).
        • Spatial + stateful → Autellix (process table) [2502.13965], TokenCake (graph API + reserved
        • pools) [2510.18586], Halo (full DAG, offline) [2509.02121].

        • Temporal + stateful → Continuum (TTL + per-tool history) [2511.02230], TokenCake's temporal half.
        • Sub-scheduler layer → KVCOMM (KV approximation) [2510.12872], RouteLLM (model choice)
        • [2406.18665].

        Paradigm claim: SMetric argues the field has been over-fitting schedulers to a false trade-off (reuse vs balance)

        carried over from chat serving, where the global tier didn't dominate. Its shift — *treat the first request as the only

        balancing decision, and let the global tier absorb the rest* — is a genuine reframing, comparable in spirit to Autellix's

        "lift LAS from per-call to per-program" [2502.13965] and Continuum's "queueing delay,

        not reload cost, is the real enemy" [2511.02230].

        Adoption evidence (strong for a research paper): SMetric is the production scheduler comparison itself — it is

        built on and evaluated against the Bailian production scheduler at "one of the world's largest token providers," on

        vLLM + LMCache + Mooncake (the de-facto open stack), and promises to open-source both code and traces

        [2607.08565]. This is materially stronger adoption grounding than the peers: Autellix,

        TokenCake, and KVCOMM are all [实现未公开] [2502.13965]

        [2510.18586] [2510.12872], whereas only Continuum

        [2511.02230] and RouteLLM [2406.18665] have public repos.

        SMetric's "at time of ingestion, promised on publication" status sits between the two.


        5. 未探索方向 — hybrid / adaptive directions from the cluster #

        1. SMetric ⊕ Continuum (spatial routing + temporal TTL). SMetric decides where the first request lands; Continuum
        2. decides how long a session's KV survives between turns [2511.02230]. A combined router would place

          for balance and TTL-pin the stuck follow-ups, so that SMetric's session_not_evicted guard fires far less often

          (fewer needless rebalances). Neither paper explores giving the stickiness decision a cost-benefit TTL.

          1. Adaptive first-request fraction under global-tier pressure. SMetric's OVERLOAD/HIT_RATIO are static
          2. [2607.08565]. Borrowing 2511.00739's throughput-gain-ratio idea [2511.00739], the router could measure the realized global-tier fetch latency

            online and dynamically shift how aggressively it balances first requests — balancing fewer of them when the global

            tier saturates (exactly the 235B regime where load-balance-only collapsed).

            1. Compaction-aware turn inference (SMetric ⊕ KVCOMM). Since message-count breaks under context compaction, KVCOMM's
            2. anchor-based offset estimation [2510.12872] could supply a content-similarity estimate of the true

              reusable prefix, making est_hit robust to compacted/forked histories — closing Attack 2.

              1. Joint CPU-GPU balanced routing (SMetric ⊕ 2511.00739). SMetric balances only GPU token load; 2511.00739 shows
              2. tool execution can dominate E2E latency [2511.00739]. A session that will issue

                heavy CPU tool calls between turns leaves its instance idle — a signal SMetric ignores. A router that co-schedules the

                expected CPU idle window (à la TokenCake's temporal offload gate [2510.18586]) with instance

                placement is an open hybrid.

                1. Two-dimensional routing: instance × model (SMetric ⊕ RouteLLM). SMetric routes to an instance assuming one model;
                2. RouteLLM routes to a model assuming one instance pool [2406.18665]. In a mixed-model agentic

                  cluster (planner=large, executor=small), a router that jointly picks (model tier, instance) while preserving

                  stateless turn inference is unexplored and directly serves the multi-model agent trend both papers gesture at.

                  1. Offline-optimal + online-stateless bridge (SMetric ⊕ Halo). Halo proves near-optimal placement is achievable when
                  2. the DAG is known [2509.02121]; SMetric proves a cheap stateless heuristic works online. Using

                    Halo-style offline placement to calibrate SMetric's guards per workload class (rather than hand-tuning

                    OVERLOAD/HIT_RATIO) could close the unquantified gap-to-optimal noted in Attack 3.