Benchmark Sync And Async Recrypt

recrypt() blocks the event loop while it re-encrypts a document. recryptAsync() does the same work on libuv's thread pool. The benchmark in packages/native-with-source/benchmarks/recrypt-server.js shows what that means for a server.

It starts an HTTP server in a child process with two routes:

• /recrypt?mode=sync or /recrypt?mode=async re-encrypts a PDF with userPassword, ownerPassword and userProtectionFlag: 4.
• /ping answers at once, like any cheap request on a busy server.

The parent process sends recrypt requests at a fixed concurrency and pings the server the whole time, so the client never shares the server's event loop.

Run It

The benchmark is a script you run by hand; it is not part of the test suite. From the root of a source checkout with a built addon:

npm run native:bench:recrypt
npm run native:bench:recrypt -- --input large.pdf --requests 32 --concurrency 8
npm run native:bench:recrypt -- --json

Inside packages/native-with-source, npm run bench:recrypt takes the same options.

Option	Default	Meaning
--input	generated PDF	The PDF to recrypt.
--copies	2	Copies of the test fixture in the generated PDF, about 8 MB each.
--requests	16	Recrypt requests per mode.
--concurrency	4	Recrypt requests in flight at once.
--ping-interval	10	Milliseconds between one ping's answer and the next ping.
--json	off	Print the measurements as JSON instead of a table.

runBenchmark(options) and formatTable(result) are exported for scripts. The file is type-checked against the package's declarations.

Read The Results

Row	Meaning
wall time	Time for all recrypt requests of the mode.
recrypts per second	Recrypt throughput.
recrypt p50, p95	How long a client waited for a recrypt response.
pings answered	Pings the server answered during the run.
ping p50, p99, max	How long a cheap request waited. This is the cost other clients pay.
loop delay	How late the server's event loop ran, from monitorEventLoopDelay. The maximum is the longest stall.
loop utilization	Share of the run the server's event loop was busy, from eventLoopUtilization.

Example Results

A 15.9 MB generated PDF, 16 requests per mode with 4 in flight, on Linux x64 with 4 CPUs, Node.js 22.22 and the default UV_THREADPOOL_SIZE of 4:

	recrypt	recryptAsync
wall time	4.55 s	4.68 s
recrypts per second	3.5	3.4
recrypt p50	1107 ms	1165 ms
recrypt p95	1667 ms	1228 ms
pings answered	9	425
ping p50	520 ms	0.6 ms
ping p99	1076 ms	2.3 ms
loop delay max	751 ms	6.5 ms
loop utilization	100%	5%

Throughput is the same, because recryptAsync runs one recrypt at a time just as the synchronous server does. The difference is everything else. With recrypt the server answered 9 pings in four and a half seconds, and a ping waited up to a second. With recryptAsync pings were answered in under a millisecond, and the event loop was idle 95% of the time.

The same run against a 95.6 MB PDF (--copies 12) took about 26 s per mode. With recrypt the server answered 10 pings, and the event loop stalled for up to 4.9 s; with recryptAsync it answered 2516 pings and never stalled for more than 16 ms.

Times depend on the machine, the document and its encryption; run the benchmark on your own documents to size a server.