4N parallel V8 isolates per Worker request.
Burn CPU across real parallel isolates on Cloudflare Workers. Hash, transform, simulate, render, evolve — work that single-threaded JavaScript can't fan out, this library can. Pick a size below and watch the topology selector decide whether to run in-DO, hybrid, or tree.
fetch / KV / AI / R2, plain Promise.all on one isolate
already gives you that for free.
When to use →
① Hero · pool.map over a CPU-bound function
view source
Each item renders one Mandelbrot tile: a horizontal slab of an image at 20,000-iteration depth around the cardioid. Roughly half a second to a few seconds of CPU per tile depending on intensity. Pick a fan-out size and an intensity preset; the library picks the topology.
Why doesn't every size beat sequential?
The library's per-call dispatch floor is roughly 50–150 ms (Durable Object RPC + leaf coordination) plus the time to ship N task envelopes across the wire. Speedup is real only when per-task CPU ≫ dispatch floor. At small N (4–16) with the medium preset, per-tile CPU is in the 200–400 ms band — comparable to dispatch — so speedup can land near 1×. At N≥64, per-task work starts dominating and speedup climbs sharply. The heavy preset crosses 5× at N=64 and 50×+ at N=512; extreme lands ~100× at N=512.
Tip: the first request after a quiet period pays a one-time
~300–400 ms DO cold-start. The library's autoWarm: true
default fires that off in parallel with the first dispatch, so
you'll see it absorbed automatically — but the cold curve is
still measurably slower than the warm curve. Click a size or
intensity tab twice to see the warm-path numbers.
await pool.map(renderMandelbrotTile, slabs)② Topology · how the selector picks
view source
The auto-selector reads items.length and picks one of three shapes.
Hover any row to see the math; click Run to fan out and update
the live numbers.
| Size | Topology | Math | V8 isolates | Live |
|---|---|---|---|---|
| 4 | in-do | 1 DO × 4 loaders | 4 | — |
| 32 | hybrid | ⌈32/4⌉=8 leaves × 4 | 32 | — |
| 128 | hybrid | ⌈128/4⌉=32 leaves × 4 | 128 (4N) | — |
| 256 | tree | K=2 tiers, F=8 | 4·F² | — |
| 512 | tree | K=3 tiers, F=8 | 4·F³ | — |
③ Every primitive, hands-on
view sourceEach card runs against the live test worker. The code shown is the actual call being made.
④ Scheduler · reactive job queue
view source
Enqueue a CPU-bound burst (each job: 1M LCG iterations). Reactive
dispatch starts work the moment a slot frees — no alarm-batched
delay. Fair round-robin across tenantId.
await scheduler.enqueue({ fn, args, tenantId, idempotencyKey })⑤ Actor · pinned state across submits
view sourceA counter Actor whose state lives in a Coordinator DO's SQLite. Each submit mutates the state in place; the runtime structured-clones it after each call. Close the Actor; the state is gone.
await actor.submit((state) => ++state.count)⑥ VM · sandboxed user code over HTTP
view source
Paste a function expression. It runs in a fresh sandboxed V8 isolate
with globalOutbound: null (no fetch) and zero bindings exposed.
Bearer-auth required. Don't try to bypass the sandbox — it's
load-bearing.
click Submit to run
⑦ Cancel · live AbortSignal
view source
Start a 1M-iteration SHA chain inside an isolate. Hit Cancel;
the request closes; env.signal.aborted trips inside the loaded
isolate; the loop returns early. Watch the iteration counter stop.
const cancel = new CancelToken();
await pool.submit(longLoop, iters, { cancel });
// hit /demo/cancel-start; close the SSE to fire cancel.⑧ Bench · live edge measurements
bench-results-live.jsonThe honest curve. Each row: Mandelbrot tile renders over N parallel V8 isolates vs the sequential per-tile baseline ×N. Speedup grows with size as the per-call dispatch floor amortizes; tree topology kicks in at 256.
| Size | Topology | Sequential | Parallel | Speedup | Run live |
|---|