Tensorlake Sharded Burst Benchmark

This repo has two Tensorlake benchmark modes. The intended distributed mode is bench:tensorlake:nested: it creates Tensorlake shard sandboxes first, then each shard sandbox creates 100 child sandboxes in parallel and runs node -v in each child.

Default scale:

• 1000 shards
• 100 sandboxes per shard
• 100000 total sandbox creations

Each result records:

• createMs: time to create the sandbox
• commandMs: time to run node -v after creation
• ttiMs: createMs + commandMs

Setup

cp env.example .env
npm install

Set TENSORLAKE_API_KEY in .env.

Run

npm run bench:tensorlake:nested

For a smaller smoke test:

npm run bench:tensorlake:nested -- --shards 2 --children-per-shard 3

Outputs are written to results/<run-id>/raw.jsonl, results/<run-id>/shards.jsonl, and results/<run-id>/summary.json.

Options

npm run bench:tensorlake:nested -- \
  --shards 1000 \
  --children-per-shard 100 \
  --shard-create-concurrency 100 \
  --shard-command-timeout-ms 300000

The older bench:tensorlake command is a flat local burst harness. It is useful for small smoke tests, but it is not the correct topology for the 1000-shard benchmark because it makes the laptop initiate every child sandbox request.

The nested worker uses Tensorlake APIs directly from inside each shard sandbox, so the laptop only creates the shard sandboxes and starts one worker command per shard.