Monitoring and performance metrics
PAI-Sandbox provides multi-level Grafana dashboards (sandbox, template, and instance) and benchmark performance data from load testing. This topic describes how to access the dashboards and the capacity and throughput you can expect in high-concurrency workloads such as RL training and agent evaluation.
Grafana dashboards
On the sandbox workspace, template details page, or sandbox details page, click Open monitoring page to open the Grafana dashboard for that scope.
The dashboards on the sandbox workspace and the template details page display the following metrics:
Sandbox:
-
Sandbox Count: total sandbox count, broken down by running and paused.
-
Sandbox Creation Latency: sandbox creation latency (max, avg, P50, P90, and more).
-
Top Sandbox Templates: the most-used templates by usage count.
-
Requested CPU Cores / Memory / Storage: total resource requests across all sandboxes.
Gateway API:
-
Sandbox Creation/Deletion/Connection QPM: queries per minute (QPM) for the sandbox creation, deletion, and connection control-plane APIs.
-
Sandbox Operation Count: cumulative count of sandbox creation, deletion, and connection operations over the selected time range.
-
Proxy Request: QPS, latency, and error rate for the Gateway proxy API that routes requests to sandbox services.
The sandbox details page dashboard shows runtime metrics for a single sandbox, including CPU and memory utilization, network, disk, and process status.
Performance metrics
PAI-Sandbox has been load-tested at large-scale concurrency. The following benchmark data helps you estimate capacity and throughput for high-concurrency workloads such as RL training and agent evaluation.
Test environment
-
Cluster: PAI MicroSandbox compute cluster.
-
Node pool size: 300 nodes.
-
ecs.c6.13xlarge(52 vCPU / 91 GiB) × 277 -
ecs.c6.26xlarge(104 vCPU / 183 GiB) × 23
-
-
Test date: April 2026.
Test methodology
A client script submits sandbox creation requests to the control plane in bulk. The test starts when the script begins sending requests and ends when the last sandbox starts. The elapsed time is the concurrent sandbox startup duration.
Template spec
Code Interpreter (1 CPU / 2 GiB memory / 8 GiB disk)
Test results
At 10,000+ concurrent sandboxes, PAI-Sandbox sustains an overall QPS above 750 with a per-node P99 startup latency under 600 ms, meeting the demands of large-scale agent evaluation and RL rollout workloads.
|
Sandboxes created |
Concurrent startup duration |
Overall QPS |
Avg node latency |
Node P50 latency |
Node P99 latency |
|
10,000 |
13.23 s |
755.74 |
396.02 ms |
395.81 ms |
502.73 ms |
|
20,000 |
26.17 s |
764.23 |
457.62 ms |
437.96 ms |
582.07 ms |
Note: The concurrent startup duration is the total time for all sandboxes to start, which includes client-side concurrency limits and other constraints. For single-sandbox startup time, refer to the node latency figures.