Performance regression after switching to v8 serialization format #10409
Description
Using the v8 serialization format rather than JSON for caches (#9971) seems to have caused a 2.33x performance regression to the resolveDependenciesTree function on v11. I noticed this when testing pnpm against a large repro I use at work and observing pnpm install consistently take much longer than it does on v10.
Measurements
I profiled the resolveDependenciesTree function before and after commit 9b344c898. Commit d392c3d05 is the parent commit of 9b344c898.
| Test | Trial 1 | Trial 2 | Trial 3 | Trial 4 | Trial 5 | Average (ms) |
|---|---|---|---|---|---|---|
| JSON (d392c3d) | 19,758.14 | 20,165.70 | 20,887.73 | 20,538.55 | 21304.28 | 20,530.88 |
| V8 buffers (9b344c8) | 46,971.59 | 47,418.23 | 47004.95 | 50838.92 | 46840.28 | 47,814.79 |
The changes I made to support profiling can be found here: a50b919
Each test run had a warmup installation and was done from a fully populated pnpm store.
Specs
MacBook Pro (16-inch, Nov 2023)
Chip: M3 Pro
Memory 36 GB
Node.js v24.11.0
macOS 26.2
CPU Profiles
Looking at a left-heavy CPU profile, the readV8FileStrictAsync and v8.deserialize calls take around ~33s.
This compares to ~7s for JSON.parse in the prior commit.
This was comparing 2 CPU profiles of the ~10 that I took.