Skip to content

Secure, Fast, and Extensible Sandbox runtime for AI agents.

14.5 KB
Raw Blame

title: Fast Sandbox Performance description: ACK measurements of the fast-sandbox integration — serial and concurrent creates, request latency, live snapshot, and same-node pause/resume.

Performance

ACK (Alibaba Cloud Container Service for Kubernetes) is Alibaba Cloud's managed Kubernetes service. This page reports measurements collected on a two-node ACK cluster on 2026-10-08: serial and concurrent creates, request latency, live snapshot, and same-node pause/resume through the OpenSandbox SDK. Runtime logs provide phase timings alongside the SDK results. These are engineering measurements, not a production latency guarantee.

The warm create path admits work into pre-warmed capacity (see Scheduling) and restores a snapshot instead of booting a kernel (see Firecracker). Pause/resume uses the submit-and-converge mutation path (see High Availability). Capacity, runtime work, and the readiness boundary still determine the observed latency and success rate.

How to read these numbers

  • SDK readiness includes the Python SDK, lifecycle server, FastPath, ingress gateway, and guest /ping 200. Runtime creation ends earlier.
  • Warm creates have template artifacts cached and Fastlet capacity prepared. First-use artifact delivery is a separate cost.
  • Sample counts and failures are reported with latency. Successful-attempt percentiles do not describe failed or timed-out requests, or prove throughput.
  • Batch differences matter: client resources and placement, and artifact-store persistence changed between the initial serial baseline and later runs. The two batches are reported separately and are not a controlled A/B comparison.

Download the raw SDK report for individual samples, failures, runtime phases, image digests, and settings.

Environment

Dimension ACK environment (2026-10-08)
Cluster Two ACK nodes, Kubernetes 1.36.2-aliyun.1, containerd 2.1.9
CPU Per node: AMD EPYC 9T95, 192 CPUs exposed, 191.45 allocatable
Memory Approximately 742 GiB allocatable per node
OS / kernel ContainerOS/Lifsea, 5.10.134-19.6.1.lifsea8.x86_64
StateRoot 50 GiB XFS reflink=1 loop filesystem per node, backed by an ext4 ESSD data disk
Pool Five Fastlets, four slots each: 20 slots
Snapshot workload Alpine 3.19 + execd, 1 vCPU, 2 GiB memory, 2 GiB rootfs
Artifact store RustFS on a persistent hostPath on one node during the later runs
Firecracker 1.16.1

The SDK deployment uses server, controller, and runtime agent release-1.1.1-rc.1. Its derived Fastlet image adds GNU coreutils cp 9.11; working reflink support requires both the filesystem and copy tool. The XFS startup rejection change was not deployed during measurement. The template artifact digest is d78575470094c6fe1a2da2a71b7578971dff9e77cd5381dee236dd53e01b54cb, cached on both nodes. SDK version: 1.1.1rc2.dev106+gc7dc78a4.d20261008.

Create

SDK creates

Each SDK create run has two untimed warmups and 100 timed attempts, with template artifacts already cached. Timing starts before Sandbox.create_from_template() and ends at guest /ping 200 through the gateway. Each sandbox is killed after readiness. Concurrent batches have a five-second reclamation pause; deletion and batch pauses are excluded from individual latency. The later serial and concurrent runs retain the same pool configuration.

The in-cluster client disables metrics, uses a ten-second HTTP request timeout, and polls readiness every 10 ms (SDK default: 200 ms). Percentiles use the nearest-rank method. Client differences between the initial serial baseline and later comparison runs are noted below.

Measurement Success p50 p90 p99 Mean
SDK create, initial serial baseline 100/100 64.0 ms 93.8 ms 108.3 ms 70.6 ms
SDK create, serial repeat alongside concurrency runs 100/100 77.9 ms 105.3 ms 112.5 ms 78.2 ms
SDK create, concurrency 10 100/100 102.7 ms 120.1 ms 125.4 ms 102.4 ms
SDK create, concurrency 20 51/100 136.7 ms 163.0 ms 197.0 ms 134.2 ms

The initial 100-create validation is retained as the warm serial baseline. Its exact p50 is 63.995 ms, mean 70.587 ms, and range 53.811–113.326 ms; all 100 creates reached guest readiness. It used the same cluster, template, pool shape, SDK version, timeout, and 10 ms readiness criterion. Its client Pod had a two-CPU limit with placement not explicitly pinned; the later comparison client has a four-CPU limit and runs on one selected node. RustFS persistence also changed between runs, although these warm creates used cached template artifacts.

The 77.9 ms repeat remains visible because it was collected alongside the concurrent runs with their client configuration. Both serial batches are in the raw report and are not pooled. Their difference is not a controlled experiment that isolates a host, client-CPU, or storage change.

Runtime logs from the same creation windows give the following breakdown. These include each run's two untimed warmups, so their sample counts differ from the SDK table. They stop at runtime creation, before SDK readiness:

Load Runtime create p50 Rootfs preparation p50 VM state load/resume p50 Runtime log samples
Serial repeat 50.0 ms 2.31 ms 0.233 ms 102
Concurrency 10 34.1 ms 2.76 ms 0.234 ms 102
Concurrency 20 33.5 ms 2.63 ms 0.225 ms 53

The initial serial validation's separate runtime window has a 39.43 ms runtime-create median and 2.286 ms rootfs-preparation median across 103 creates (100 timed, two warmups, and one subsequent lifecycle create).

::: warning Saturation result The concurrency-20 latency distribution contains successful attempts only. There were 48 SandboxRateLimitException failures (HTTP 429 / Fastlet CapacityRejected) and one readiness timeout with repeated gateway HTTP 503. Four test CRs also stuck in deletion with an assignment annotation/status projection conflict. Their stale placement projections were cleared to let the controller perform cleanup before the next experiment. This run does not establish reliable 20-way admission or a sustainable requests-per-second rate. The implementation cause of the admission/projection failure needs a separate fix; replacing the host disk alone does not resolve it. :::

Request latency

The SDK samples use one warm sandbox and 100 sequential calls per path:

Path Success p50 p90 p99 Mean
Warm SDK /ping through the gateway 100/100 0.92 ms 1.15 ms 5.91 ms 1.07 ms
SDK commands.run('echo ack-perf'), through completion and stream EOF 100/100 202.8 ms 204.3 ms 207.1 ms 203.2 ms

The execd command handler has a default 200 ms stream-close grace period; the observed approximately 203 ms is consistent with that behavior. It is not a measurement of shell process startup alone.

Live snapshot

One live snapshot was measured:

Phase Measured
Snapshot request accepted 14.4 ms
Request → available snapshot (Ready in the SDK) 60.63 s
VM frozen during dump 11.14 s
Runtime artifact publication 46.22 s
Artifact set 2 GiB rootfs + 2 GiB memory

The SDK source was healthy afterward, with its file marker and boot ID preserved. Requests were not sampled throughout its freeze, so this establishes no zero-interruption claim. The initial probe used an incorrect terminal-state check and is excluded from the timings.

Restore from snapshot

Creating a new sandbox from a published live snapshot was not measured in this ACK run. Cached checkpoint restoration was measured as part of same-node pause/resume; template restoration is covered by Create. These operations have separate completion boundaries.

Pause / resume

For this test only, the idle pool was restricted to one existing node, retaining five Fastlets and four slots per Fastlet. Every sample's source and destination Kubernetes node were verified through the CR's placement and the Fastlet Pod's spec.nodeName. All three resumes moved to a different Fastlet on that same node and used the node's checkpoint cache. The original pool selector was restored afterward.

Measurement Median Observed range (3 samples)
SDK pause request accepted 22.4 ms 22.0–25.5 ms
Pause request → durable Paused, including upload and release 61.14 s 61.13–61.23 s
VM frozen during checkpoint dump 11.15 s 11.13–11.18 s
Publish approximately 4 GiB (2 GiB rootfs + 2 GiB memory) 45.61 s 45.61–45.63 s
Resume runtime creation, after cached artifacts are available 35.49 ms 33.64–35.50 ms
Resume rootfs preparation 2.92 ms 2.64–3.26 ms
SDK same-node resume → guest /ping 200 62.66 s 58.29–65.16 s

All three samples preserved the filesystem marker, tmpfs marker, boot ID, background PID, and process start time. The sandbox identity also stayed the same. No checkpoint artifact-delivery attempt appears in the corresponding Fastlet logs. Rootfs checkpoint cloning took 0.88–0.94 ms with zero full-copy time. Three samples establish this observed range, not a reliable p99.

::: warning Same-node resume remains slow Cached runtime restoration is tens of milliseconds, but usable SDK resume is approximately a minute in this deployment. Runtime logs publish the route immediately after restoration, leaving most elapsed time in endpoint/guest availability and client waiting. These measurements do not isolate which network, gateway, endpoint-resolution, or retry mechanism causes that gap. The gap cannot be explained by copying a 2 GiB rootfs or fetching a checkpoint onto another node. Do not substitute the runtime number for SDK resume.

The probe used a four-minute resume budget and ten-second HTTP request timeout; the SDK's default resume budget is 30 seconds. The reported result includes actual readiness waiting and does not skip health checks. :::

Artifact delivery (DART P2P)

Not verified in this ACK run. Template artifacts were already cached on both nodes, and same-node resumes used the checkpoint cache. Cold delivery latency, origin/peer block counts, and P2P efficiency require separate measurements.

Boot characteristics

Cold kernel boot was not verified in this ACK run. Template creates restore vmstate.snap instead of booting a kernel; their measured VM state load/resume phases appear under Create.

Interpreting the results

XFS reflink support and GNU cp keep warm rootfs preparation in the millisecond range: the initial serial runtime window measured a 2.286 ms median. This measures the storage/copy phase, not total SDK readiness.

Concurrency changes admission contention; durable pause includes dumping, hashing, and uploading a checkpoint; SDK readiness includes endpoint resolution, HTTP timeouts, and retries. The saturation failures under Create are control-plane evidence, not a host-only explanation. The same-node resume gap also remains unresolved. A warm /ping, completed command stream, runtime restore, and SDK resume have different completion boundaries.

Reproducing

SDK measurements

Use a source-built SDK matching the recorded version. Set DOMAIN, API_KEY, and TEMPLATE_ID for the target cluster without putting credentials in the report. Run from an in-cluster client to retain this network boundary:

REQUEST_TIMEOUT=10 DISABLE_METRICS=1 PING_INTERVAL_MS=10 READY_TIMEOUT_WARM=180 \
N=100 CONCURRENCY=1 BATCH_PAUSE=0 REPORT_PATH=serial.json \
  python scripts/fast-sandbox-env/create-bench.py

REQUEST_TIMEOUT=10 DISABLE_METRICS=1 PING_INTERVAL_MS=10 READY_TIMEOUT_WARM=180 \
N=100 CONCURRENCY=10 BATCH_PAUSE=5 REPORT_PATH=concurrent10.json \
  python scripts/fast-sandbox-env/create-bench.py

# Saturation diagnostic: record failures as well as successful latencies.
REQUEST_TIMEOUT=10 DISABLE_METRICS=1 PING_INTERVAL_MS=10 READY_TIMEOUT_WARM=180 \
N=100 CONCURRENCY=20 BATCH_PAUSE=5 REPORT_PATH=concurrent20.json \
  python scripts/fast-sandbox-env/create-bench.py

REPORT_PATH saves raw attempts, settings, success/failure counts, and nearest-rank percentiles. A run with failed creates returns a nonzero exit status. Save the deployed image digests, host/storage details, template digest, pool shape, and timestamped Fastlet/controller logs alongside it. The SDK run used an equivalent probe that also collected lifecycle and request samples; the report retains those original samples.

For the other measurements, retain the same SDK connection settings:

  1. On one warm sandbox, time 100 sequential sandbox.is_healthy() calls, asserting success; then time 100 sandbox.commands.run('echo ack-perf') calls through stream EOF, asserting the expected output.
  2. On an idle pool restricted to one node, create a fresh sandbox for each of three lifecycle samples. Before pausing, wait for CR status.runtime.state == Ready; the SDK's early readiness result can precede that projection. Write markers to the filesystem and tmpfs and record the boot ID and a background process's PID/start time.
  3. Start the pause timer before sandbox.pause(). Poll manager.get_sandbox_info(id).status.state == 'Paused' every 200 ms; an accepted pause response alone is not a durable checkpoint.
  4. Time Sandbox.resume(id, resume_timeout=timedelta(minutes=4), health_check_polling_interval=timedelta(milliseconds=10), connection_config=config) through its normal readiness check. Verify placement on the original node and all guest markers before killing the sandbox. Restore the pool's original selector and wait for five warm, ready Fastlets.
  5. For live snapshot, time sandbox.create_snapshot() acceptance separately from manager.get_snapshot(id).status.state == 'Ready' (200 ms polling). The public snapshot API uses Ready; template builds use Succeeded. Verify the source and delete the test snapshot and sandbox.

Capture runtime dump, publish, restore, and route-publication log timestamps to distinguish checkpoint I/O from later SDK readiness. Clean up only the artifacts created by the test after their sandbox/snapshot resources are gone.

Keep hardware, kernel, Firecracker version, template digest, pool shape, client placement, and artifact-store placement fixed when comparing load shapes.