weili-0234/fs-exp-trace-3

FlashState experiment traces, part 3

Raw client-side request and response dumps of the store-policy study of 2026-11-03 (FlashState with a Mooncake host store) and of the comparison of 2026-11-04 (FlashState's full stack in the release engine against official vLLM, both with Mooncake). Each file is aiperf's raw export of one engine (one arm) over one run. Parts 1 and 2 (weili-0234/fs-exp-trace and weili-0234/fs-exp-trace-2) hold the earlier campaign (from 2026-10-28) and describe the trace replay in more detail.

Study

The study asks how the host store should keep the states of a hybrid model (attention KV and recurrent checkpoints) while agent sessions wait for tool calls. The arms differ in two things:

1. Tail write-through: the state at the exact end of each request is written to the host store as soon as it exists, so that the next turn of the session can restore it instead of recomputing the tail.
2. Host-store eviction: the stock Mooncake master, or a master that evicts whole recovery units (the KV blocks plus the recurrent checkpoint needed to resume from one point) by GreedyDual-Size.

arm (name in CELLS.md)	engine	tail write-through	host-store eviction
rel-gds-mc512-hwm98	release	no	stock master (the baseline)
rel-wt-mc512-hwm98	release	yes	stock master
rel-wt-units-mc512-hwm98	release	yes	recovery units; leaf units evicted last; a session's older tail units marked dead
rel-wt-units-keep-flat-mc512-hwm98	release	yes	recovery units; one tier (plain GreedyDual-Size); no dead marks
fs-mc512-wt-units-hwm98	research	yes	recovery units; leaf units evicted last; older tail units marked dead
fs-mc512-wt-units-keep-hwm98	research	yes	recovery units; leaf units evicted last; no dead marks
fs-mc512-wt-units-keep-flat-hwm98	research	yes	recovery units; one tier; no dead marks

The release engine keeps recurrent checkpoints at 4,352-token boundaries (build b39b47c0, on release cb8036cc); the research engine matches prefixes in units of 1,088 tokens (build 06ec402e).

Comparison of 2026-11-04 (round fsr1104)

FlashState's full stack in the release engine (build 2c0fe4a6) against official vLLM (ec8a3725, no file changed), both with a Mooncake host store. In each cell both arms use the same Mooncake client: the client as shipped in the image, or the copy fix (a generic client patch: the store's GPU copies run on a non-blocking CUDA stream through pinned staging buffers instead of synchronously on the default stream).

arm	engine	Mooncake client	host-store eviction
fsr-final-mcfix-mc512-hwm98	release, full stack	copy fix	recovery units (GreedyDual-Size, anchor tier, idle demotion)
official-mcfix-mc512-hwm98	official	copy fix	stock master
fsr-final-wheel-mc512-hwm98	release, full stack	as shipped	recovery units (GreedyDual-Size, anchor tier, idle demotion)
official-mc512-hwm98	official	as shipped	stock master

The full stack stores recurrent checkpoints at the frontier (with prefill export), evicts GPU prefix-cache entries by GreedyDual-Size with leaf units last, writes the exact tail state of each request through to Mooncake, and looks up the newest stored state. Both arms match prefixes every 1,088 tokens (--prefix-match-unit 1088).

Setup (all cells)

• Model: GLM-5.3-Flash NVFP4, 1M context, thinking kept. Output lengths are fixed by the trace (ignore_eos).
• Workload: AgentX agentic sessions replayed live with token IDs (the model's own outputs are fed back into the next turn); 60, 64, 68, 72, 76 or 80 concurrent sessions.
• Hardware: NVIDIA B200, tensor parallel 2 per engine, two engines (two arms) side by side on one 4-GPU node, so both arms of a cell run at the same time on the same node.
• GPU KV pool 66 GB per GPU; Mooncake host store 512 GiB per engine, eviction high watermark 0.98.
• About 13 minutes of warm-up, then 2,700 s of profiling (metadata.benchmark_phase tells them apart). The study's numbers use the requests that start in [900, 2700) s of profiling.

Layout

ci/<round>/<cell>-<slot>/client/aiperf_artifacts/profile_export_raw.jsonl.gz
CELLS.md

• ci/rel1103/: release engine. Cells q64r* run at 64 sessions; cells s<sessions>a and s<sessions>b run the other loads (b repeats a with the two slots swapped).
• ci/units1103/: research engine at 64 sessions (cells k64*).
• ci/fsr1104/: cells xm<sessions>a|b pair the two copy-fix arms and cells xw<sessions>a|b the two arms on the client as shipped; b repeats a with the two slots swapped. A trailing digit (for example xm80a3) marks a rerun of a cell whose node the cluster took back before it finished; only finished runs are here.
• <slot> is g01 (GPUs 0-1) or g23 (GPUs 2-3).
• CELLS.md lists every folder with its sessions, arm and the sha256 of its file.

Each file has one JSON line per request: metadata (session number, conversation and turn, request timestamps in ns, warm-up or profiling phase), payload (the prompt as token IDs and the sampling parameters), responses (the streamed packets with their timestamps, including the returned token IDs), request_headers and status.

Checks

Every file is checked against the sha256 recorded when its cell was collected, and passes a secret scan before upload (credentials, service URLs, private network addresses, home-directory paths). A match only inside the model's own generated text does not count.

Access is gated: request access on this page; each request is approved by hand.