0
0
mirror of https://github.com/nodejs/node.git synced 2024-11-30 07:27:22 +01:00
nodejs/benchmark/fs/readfile-partitioned.js
Jamie Davis 67a4ce1c6e
fs: partition readFile against pool exhaustion
Problem:

Node implements fs.readFile as:
- a call to stat, then
- a C++ -> libuv request to read the entire file using the stat size

Why is this bad?
The effect is to place on the libuv threadpool a potentially-large
read request, occupying the libuv thread until it completes.
While readFile certainly requires buffering the entire file contents,
it can partition the read into smaller buffers
(as is done on other read paths)
along the way to avoid threadpool exhaustion.

If the file is relatively large or stored on a slow medium, reading
the entire file in one shot seems particularly harmful,
and presents a possible DoS vector.

Solution:

Partition the read into multiple smaller requests.

Considerations:

1. Correctness

I don't think partitioning the read like this raises
any additional risk of read-write races on the FS.
If the application is concurrently readFile'ing and modifying the file,
it will already see funny behavior. Though libuv uses preadv where
available, this doesn't guarantee read atomicity in the presence of
concurrent writes.

2. Performance

Downside: Partitioning means that a single large readFile will
  require into many "out and back" requests to libuv,
  introducing overhead.
Upside: In between each "out and back", other work pending on the
  threadpool can take a turn.

In short, although partitioning will slow down a large request,
it will lead to better throughput if the threadpool is handling
more than one type of request.

Fixes: https://github.com/nodejs/node/issues/17047

PR-URL: https://github.com/nodejs/node/pull/17054
Reviewed-By: Benjamin Gruenbaum <benjamingr@gmail.com>
Reviewed-By: Tiancheng "Timothy" Gu <timothygu99@gmail.com>
Reviewed-By: Gireesh Punathil <gpunathi@in.ibm.com>
Reviewed-By: James M Snell <jasnell@gmail.com>
Reviewed-By: Matteo Collina <matteo.collina@gmail.com>
Reviewed-By: Sakthipriyan Vairamani <thechargingvolcano@gmail.com>
Reviewed-By: Ruben Bridgewater <ruben@bridgewater.de>
2018-02-01 11:11:48 +01:00

87 lines
1.9 KiB
JavaScript

// Submit a mix of short and long jobs to the threadpool.
// Report total job throughput.
// If we partition the long job, overall job throughput goes up significantly.
// However, this comes at the cost of the long job throughput.
//
// Short jobs: small zip jobs.
// Long jobs: fs.readFile on a large file.
'use strict';
const path = require('path');
const common = require('../common.js');
const filename = path.resolve(__dirname,
`.removeme-benchmark-garbage-${process.pid}`);
const fs = require('fs');
const zlib = require('zlib');
const assert = require('assert');
const bench = common.createBenchmark(main, {
dur: [5],
len: [1024, 16 * 1024 * 1024],
concurrent: [1, 10]
});
function main(conf) {
const len = +conf.len;
try { fs.unlinkSync(filename); } catch (e) {}
var data = Buffer.alloc(len, 'x');
fs.writeFileSync(filename, data);
data = null;
var zipData = Buffer.alloc(1024, 'a');
var reads = 0;
var zips = 0;
var benchEnded = false;
bench.start();
setTimeout(function() {
const totalOps = reads + zips;
benchEnded = true;
bench.end(totalOps);
try { fs.unlinkSync(filename); } catch (e) {}
}, +conf.dur * 1000);
function read() {
fs.readFile(filename, afterRead);
}
function afterRead(er, data) {
if (er) {
if (er.code === 'ENOENT') {
// Only OK if unlinked by the timer from main.
assert.ok(benchEnded);
return;
}
throw er;
}
if (data.length !== len)
throw new Error('wrong number of bytes returned');
reads++;
if (!benchEnded)
read();
}
function zip() {
zlib.deflate(zipData, afterZip);
}
function afterZip(er, data) {
if (er)
throw er;
zips++;
if (!benchEnded)
zip();
}
// Start reads
var cur = +conf.concurrent;
while (cur--) read();
// Start a competing zip
zip();
}