Skip to content

Io\Poll\Context::wait() is O(n) in the number of level-triggered Read watchers #25602

Open

Description

@medilies

Description

Since #24017 (fix for GH-23948), Io\Poll\Context::wait() takes time proportional to the number of registered level-triggered Read watchers, even when none of them is ready. Before each backend wait it iterates over every watcher of the context to look for data in the stream read buffer:

/* Buffered watchers go first, then a non-blocking backend check */
php_io_poll_watcher_object **buffered = NULL;
int n_buffered = 0;
if (intern->buffered_readers) {
ZEND_HASH_FOREACH_VAL(intern->watchers, zval *zv) {
php_io_poll_watcher_object *watcher = PHP_POLL_WATCHER_OBJ_FROM_ZOBJ(Z_OBJ_P(zv));
if (watcher->disarmed || !php_io_poll_watcher_reads_buffered(watcher)
|| !watcher->ext_ops->has_buffered_data(watcher->handle)) {
continue;
}
if (!buffered) {
buffered = safe_emalloc((size_t) max_events, sizeof(*buffered), 0);
}
buffered[n_buffered++] = watcher;
if (n_buffered == max_events) {
break;
}
} ZEND_HASH_FOREACH_END();

The following code:

<?php

use Io\Poll\{Backend, Context, Event};
use Time\Duration;

foreach ([100, 1000, 10000, 20000] as $n) {
    foreach (['level' => [Event::Read], 'edge' => [Event::Read, Event::EdgeTriggered]] as $mode => $events) {
        $ctx = new Context(Backend::Epoll);
        $pairs = [];
        for ($i = 0; $i < $n; $i++) {
            $pairs[] = $pair = stream_socket_pair(STREAM_PF_UNIX, STREAM_SOCK_STREAM, STREAM_IPPROTO_IP);
            $ctx->add(new StreamPollHandle($pair[0]), $events);
        }

        $zero = Duration::fromSeconds(0);
        $start = hrtime(true);
        for ($i = 0; $i < 5000; $i++) {
            $ctx->wait($zero);
        }
        printf("%-5s n=%-6d %8.2f us per wait()\n", $mode, $n, (hrtime(true) - $start) / 5000 / 1e3);

        unset($ctx);
        foreach ($pairs as [$a, $b]) {
            fclose($a);
            fclose($b);
        }
    }
}

Resulted in this output (needs ulimit -n above 40000):

level n=100        1.19 us per wait()
edge  n=100        0.38 us per wait()
level n=1000       8.77 us per wait()
edge  n=1000       0.37 us per wait()
level n=10000    295.23 us per wait()
edge  n=10000      0.36 us per wait()
level n=20000    933.23 us per wait()
edge  n=20000      0.40 us per wait()

Expected: the level rows stay close to the edge rows, since no handle is ready and the epoll backend is O(ready events). maxEvents makes no difference.

Impact on an event loop: a TCP echo server (32 active ping-pong connections plus N idle ones, loopback, 4 CPUs) on revolt/event-loop 1.1.0-RC1 IoPollDriver drops from about 100k round trips/s with 1,000 idle connections to 56k with 10,000 and 27k with 20,000. Server CPU per round trip goes from 9.9 µs to 37.5 µs. The same server on EvDriver (ext-ev) stays at about 105k round trips/s and 9.4 µs at every size. Level-triggered Read is the default and what the event loop drivers use, so they all pay this cost on every tick.

A possible direction that keeps the GH-23948 behaviour without breaking ABI: a stream's read buffer only grows in php_stream_fill_read_buffer(), and php_stream already has poll_watchers (used by php_io_poll_stream_notify_close()). That function could notify the stream's watchers when the buffer is non-empty, queueing them on their context. wait() would then check only the queued watchers, dropping those whose buffer has been drained, instead of every watcher.

PHP Version

PHP 8.6.0 (cli) (built: Nov 19 2026 15:18:16) (ZTS)

Official php:8.6-cli-alpine Docker image.

Operating System

Alpine Linux 3.24 (Docker), host kernel Linux 7.0 x86_64

Activity

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions