bench: add generated large App Router benchmark for Turbopack scaling (#101502)
### What?
This adds `bench/large-app-router`, a large generated App Router app. It
is meant to be the shared target for measuring how Turbopack `next
build` / `next dev` scale with the number of CPU cores (see #100250).
- **`scripts/generate.mjs`**: a deterministic generator. A seeded PRNG
gives byte-identical output for the same args. It takes `--routes`,
`--components` and `--seed`. With the defaults (`--routes 5000
--components 3000`) it writes ~21k files, all gitignored:
- 100 sections, each with a layout, `loading.tsx`, `error.tsx` and an
index page
- 5000 pages: ~10% are `[slug]` segments with `generateStaticParams`,
~15% have nested layouts
- a 3000-component shared library, ~35% of it `'use client'`, layered
like a design system (primitives → molecules → organisms → templates)
- a CSS Module next to every component, layout and page, plus a global
stylesheet
- imports from `lodash-es`, `lucide-react`, `clsx`, `dayjs` and `zod`
(versions already in the lockfile)
- `next/dynamic`, plus `import()` in client event handlers
- 500 `app/api/*/route.ts` handlers
- **`scripts/bench-build.mjs`**: runs N cold `next build`s. It reports
per-run wall time and the Turbopack compile time (the `run-turbopack`
span from `.next/trace`), plus mean ± sample stddev, with optional JSON
output. `--cpus <n>` / `--cpu-list <list>` pin the build via `taskset`,
choosing CPUs from the process's allowed affinity.
- **README**: covers generate / build / dev / benchmark, using a release
next-swc binary, and thread-count control:
- `taskset` is recommended, because it limits Node workers too.
- `TURBO_TASKS_AVAILABLE_PARALLELISM` sets the Turbopack/tokio thread
count.
- `TOKIO_WORKER_THREADS` has no effect, because next-swc sets the tokio
worker count explicitly.
### Why?
The existing `bench/` apps are either small, purely synthetic, or
single-purpose. For example, `nested-deps-app-router-many-pages` is 1000
identical template copies with no CSS, route handlers, dynamic imports
or npm deps. Scaling work needs one realistic, reproducible,
large-enough target.
The library is layered on purpose so that every module's transitive
closure stays bounded. An unbounded random import graph makes per-page
closures grow with the library size, so the total work grows
quadratically. Doubling the size then made the compile ~9× slower, which
measures the generator rather than the bundler. With the layering, build
time grows linearly with app size: 1500 → 3000 routes went from 17.6s →
36.8s of compile at 32 CPUs.
### Initial scaling numbers
- **Source:** canary `4d51724b`. **Native binary:** the published
release `@next/swc-linux-x64-gnu@16.5.0-canary.46`, loaded via
`NEXT_TEST_NATIVE_DIR`.
- **App:** default generated size, cold builds with the default
persistent cache, 3 runs each, pinned via `--cpus`.
> **Topology caveat:** the machine has **16 physical cores / 32 logical
CPUs** (SMT, with adjacent siblings). The columns below are *logical*
CPUs, so 32 = 16 cores + HT and 16 = 8 cores + HT. This is not yet a
curve for 32 physical cores.
| Logical CPUs (mask) | `next build` wall | Turbopack compile
(`run-turbopack`) |
| ------------------- | ----------------: |
----------------------------------: |
| 8 (`0-7`) | 168.00s ± 5.38s | 121.98s ± 4.44s |
| 16 (`0-15`) | 124.33s ± 2.68s | 89.80s ± 2.75s |
| 32 (`0-31`) | 95.34s ± 0.98s | 57.56s ± 0.47s |
Speedup of the Turbopack compile: 8 → 16 logical CPUs is 1.36×, 16 → 32
is 1.56×. The parallel efficiency from 8 to 32 is ~53%.
Diagnostic: spreading 8 / 16 CPUs across one thread per physical core
(`--cpu-list 0,2,4,…`) gave 166.84s ± 6.22s / 121.88s ± 1.93s wall. That
is about the same as the SMT-packed masks above, so on this machine
adding threads helps about as much as adding physical cores. This was
measured on generated output that differs only in the section navigation
links.
### How to run
```bash
pnpm install && pnpm build
cd bench/large-app-router
pnpm prepare-bench # or: node scripts/generate.mjs --routes 500 --components 300
for c in 8 16 32; do pnpm bench-build --runs 3 --cpus $c; done
```
### Verification
- The generator is deterministic: two default runs gave the same content
hash, a different `--seed` changes the output, and `--routes 100
--components 60` produces 428 files.
- Every generated section link resolves to a page or a generated
`[slug]` param (5000/5000).
- Turbopack `next build` of the default app succeeds on every run above.
- `--cpus 4` was confirmed in `/proc/<pid>/status`: the `next-build`
process had `Cpus_allowed_list` `0-3`. Invalid, reversed or unavailable
CPU lists exit with a clear error.
- Prettier and eslint are clean. The lockfile only gains the new
importer entry.
<!-- NEXT_JS_LLM -->
<!-- fleet 844a9209-a191-4c7b-904b-5f0ad1e6b91d -->
Co-authored-by: vercel-fleet-prod[bot] <318278635+vercel-fleet-prod[bot]@users.noreply.github.com>
Co-authored-by: Tobias Koppers <1365881+sokra@users.noreply.github.com>