Mocker Trace Replay
Replay Mooncake-style traces through the mocker in offline or online mode
This guide covers trace replay support for Mooncake-style JSONL traces via python -m dynamo.replay,
which prints an AIPerf-style summary table, writes the full replay report JSON to disk, and exposes
offline|online, round_robin|kv_router, arrival_speedup_ratio, closed-loop concurrency, and
synthetic workload inputs directly.
Unlike normal dynamo.mocker usage, offline replay does not launch workers, register endpoints, or
require NATS, etcd, or a frontend. Online replay does exercise the live mock-worker runtime path.
Use this when you want to:
- benchmark scheduler behavior from a saved trace
- compare timing and cache behavior across mocker configurations
- validate replay logic in CI without bringing up a distributed stack
Harness Overview
The replay harness wires a load driver (trace file or synthetic workload generator) into one or more mocker engine simulations and tees request/token timing into a trace collector.
The load driver is either a Mooncake-style JSONL trace (timestamps, ISL/OSL, hash_ids) or a synthetic generator parameterized by isl/osl/concurrency. Single-engine simulation (SES) is the fast path for num_workers == 1 with the vLLM engine; multi-engine simulation (MES) covers aggregated multi-worker replay, disaggregated prefill/decode replay, and KV-router replay. The trace collector produces the AIPerf-style summary table, the JSON report, and the per-request timing fields consumed by downstream analysis.
Each simulation composes a different set of components. SES drives the engine core directly (scheduler + forward-pass modeling). MES composes multiple engine cores with KV transfer/offloading, KV routing, and planner simulation layered on top:
See lib/mocker/src/replay/offline/README.md for offline-harness internals (logical clock, event queue, worker model) and docs/mocker/mocker.md for engine-core details (scheduler, KV block manager).
Quick Start
Run offline replay through the dedicated replay CLI:
Run synthetic replay through the same CLI when you want fixed request shapes without a trace file:
Run synthetic workload replay when you want shared-prefix or multi-turn structure without a trace file:
python -m dynamo.replay prints an AIPerf-style summary table to stdout and writes the full replay
report JSON to disk.
Input Format
The trace file must be Mooncake-style JSONL. Each line should contain:
timestamporcreated_timeinput_lengthorinput_tokensoutput_lengthoroutput_tokenshash_ids
Example:
Replay also supports multi-turn sessions. Use the same session_id on all turns in a session. The
first turn uses timestamp or created_time; later turns may use either:
delayordelay_msdirectly- or an absolute later
timestamp, in which case replay infers the inter-turn delay from the previous turn timestamp
Example:
Replay uses two different block-size concepts for trace files:
--trace-block-size: how many tokens eachhash_idin the dataset represents- engine
block_size: the block size used by the replay engine and router when they re-chunk the synthesized tokens into sequence hashes
Public Mooncake/toolagent traces use 512 tokens per hash_id, so replaying them should normally
use --trace-block-size 512. The engine block_size can still be smaller, for example the live
vLLM benchmark setup uses block_size=64. For engine_type=sglang, replay still uses canonical
block_size internally; sglang.page_size is accepted as a compatibility alias and is normalized
into block_size before replay starts.
Replay Surfaces
python -m dynamo.replay
The dedicated replay CLI exposes:
- either a positional
trace_file, or all of--input-tokens,--output-tokens, and--request-count --replay-mode offline|online--router-mode round_robin|kv_router--num-workers--num-prefill-workers--num-decode-workers--replay-concurrency--arrival-interval-ms--arrival-speedup-ratio--trace-block-size--turns-per-session--shared-prefix-ratio--num-prefix-groups--inter-turn-delay-ms--extra-engine-args(JSON string)--prefill-engine-args(JSON string)--decode-engine-args(JSON string)--router-config(JSON string)--aic-backend--aic-system--aic-backend-version--aic-tp-size--aic-model-path--report-json
Defaults:
--replay-mode offline--router-mode round_robin
Example:
SGLang replay uses the same CLI surface. A minimal extra-engine-args file can use either
block_size directly or the compatibility alias sglang.page_size:
Both --extra-engine-args and --router-config accept partial JSON objects. Engine settings such
as block_size, engine_type, dp_size, speedup_ratio, and decode_speedup_ratio belong in
--extra-engine-args, not as top-level replay CLI flags. --trace-block-size is separate and is
used only for trace-file replay. Unspecified fields fall back to the same defaults used by
MockEngineArgs::default() and KvRouterConfig::default().
Replay has two independent AIC surfaces:
- engine timing AIC via
--extra-engine-args/ staged engine JSON - router-side prompt-load AIC via top-level
--aic-*flags together withrouter_prefill_load_model: "aic"in--router-config
Offline disagg replay uses staged engine args instead of --extra-engine-args:
--prefill-engine-argsfor the prefill worker config--decode-engine-argsfor the decode worker config--num-prefill-workersand--num-decode-workersfor pool sizes
For offline disagg replay, the staged JSON must set worker_type explicitly:
--prefill-engine-argsmust useworker_type: "prefill"--decode-engine-argsmust useworker_type: "decode"
The staged configs must also use the same engine block_size. --trace-block-size remains a
separate trace-file input knob.
Synthetic Replay
Synthetic replay bypasses trace loading and generates in-memory requests with fixed input/output lengths and optional synthetic arrival spacing:
This is useful for parameter sweeps where Mooncake-style prefix structure is not required.
When --turns-per-session > 1, --request-count is interpreted as the number of sessions rather
than the total number of emitted turns. The total completed request count becomes:
request_count * turns_per_session
Synthetic workload options:
--turns-per-session: number of turns in each synthetic session--shared-prefix-ratio: fraction of prompt blocks shared inside a prefix group--num-prefix-groups: number of shared-prefix groups;0disables grouping--inter-turn-delay-ms: constant delay applied after each completed turn before the next turn in the same session becomes eligible
Modes
Fixed-Schedule Replay
Default trace replay preserves the timestamps from the trace and simulates arrivals according to those timestamps:
This is the right mode when you want deterministic replay of the original arrival pattern.
Closed-Loop Concurrency Replay
Use --replay-concurrency to ignore first-turn trace arrival timing and keep a fixed number of
requests in flight:
This mode is useful when you want to compare scheduler behavior under a fixed offered concurrency rather than the original trace schedule.
For multi-turn sessions, concurrency mode still enforces session order and inter-turn delays:
- first-turn timestamps are ignored
- turn
n+1is not eligible until turnncompletes delay/delay_ms/ synthetic--inter-turn-delay-msare still applied after completion- TTFT is measured from actual dispatch under the cap, not from the ignored trace timestamp
Online Replay
Online replay launches the mock workers and replays the trace against the live runtime path. This is useful when you want the replay to include live request dispatch, live output handling, and the same async KV-event propagation model used by the current router integration.
Arrival Speedup
Use --arrival-speedup-ratio to compress or stretch the trace arrival process without changing the
mocker compute model. Larger values make arrivals happen sooner relative to the original trace.
Router Modes
Replay currently supports:
round_robinkv_router
kv_router uses the shared local scheduler and an in-process KV indexer. Router policy tuning is
provided through --router-config, not a dedicated top-level replay flag. In offline replay:
kv_routeris supported only whennum_workers > 1- router queueing is enabled and uses simulation time rather than wall-clock time
- KV visibility is delayed slightly relative to request lifecycle events
- queue admission is driven by router lifecycle edges (
add_request,mark_prefill_completed, andfree) - transient in-pass prefill occupancy is still approximated at the router level rather than modeled exactly
- when
router_prefill_load_modelis"aic", replay predicts one expected prefill duration per admitted request and decays only the oldest active prefill request on each worker
To compare queue policies manually, keep the same trace and engine args fixed and swap only
router_queue_policy inside --router-config:
lcfs is intentionally a worse comparison policy under saturation; use it for experiments, not as
an expected production default.
To enable router-side AIC prefill-load modeling during replay:
For offline disagg replay, the same top-level --aic-* flags are supported, but the estimator is
applied only to the prefill-stage router.
Output
The report contains:
- request counts
- input and output token totals
- virtual duration and wall-clock runtime
- request and token throughput
- prefix cache reuse ratio
- TTFT, TTST, TPOT, ITL, and end-to-end latency summaries
- output-token-throughput-per-user summaries
The dedicated replay CLI returns the same report schema as the Python APIs
dynamo.replay.run_trace_replay(...) and dynamo.replay.run_synthetic_trace_replay(...).
If --report-json is not provided, python -m dynamo.replay writes a timestamped
dynamo_replay_report_*.json file in the current working directory.
Replay Constraints
Shared replay constraints:
extra_engine_args.engine_typemust bevllmorsglang- aggregated replay requires the existing aggregated args path
- disagg replay requires both
prefill_engine_argsanddecode_engine_args - disagg replay requires
router_mode=kv_router - replay
dp_sizemust be1 - disagg replay requires matching
block_sizeinprefill_engine_argsanddecode_engine_args
Additional offline constraints:
- offline
kv_routerrequiresnum_workers > 1 - single-worker offline replay is still a dedicated fast path for
vllm, but it now supports both flat request replay and workload-driven multi-turn replay sglangstill goes through the shared multi-worker replay runtime even whennum_workers=1- offline disagg replay is a separate two-stage runtime with prefill and decode worker pools
Additional online constraints:
- the current live replay path is also limited to aggregated workers
If you violate those constraints, replay fails immediately with a validation error.
Practical Notes
python -m dynamo.replayrequires exactly one of: either a trace file, or all of--input-tokens,--output-tokens, and--request-count--replay-concurrencyworks with both trace replay and synthetic replay- mocker compute-speed knobs such as
speedup_ratiostill affect simulated timing when passed via the engine-args JSON for the chosen replay mode --arrival-speedup-ratioaffects trace timestamps, not worker compute speed--trace-block-sizeaffects only how tracehash_idsexpand into tokens--arrival-interval-msonly applies to synthetic replay--turns-per-session,--shared-prefix-ratio,--num-prefix-groups, and--inter-turn-delay-msonly apply to synthetic replay--extra-engine-args,--prefill-engine-args,--decode-engine-args, and--router-configare JSON strings on the standalone replay CLI- top-level
--aic-*flags are used only for router-side prompt-load modeling; engine timing AIC still belongs in the engine-args JSON - offline replay does not need planner runtime setup, router registration, or external event transport
- trace-file replay can use different values for
--trace-block-sizeand engineblock_size - Mooncake/toolagent traces typically use
--trace-block-size 512, while engineblock_sizeoften stays64
When To Use This vs AIPerf
Use offline replay when:
- you want a fast scheduler-only simulation
- you want deterministic CI coverage of replay behavior
- you do not need HTTP serving, frontend behavior, or network effects
Use Dynamo Benchmarking when:
- you want end-to-end benchmarking against a live endpoint
- you need frontend, transport, or cluster-level behavior
- you want AIPerf dashboards and endpoint-facing metrics