Issue with performance and compilation

[ImmanuelSamuel]

Background

I am evaluating the performance of the http_server example, comparing it to bun, which also leverages uWebSockets->uSockets under the hood (I think).

Observed Behavior:

• For ~1,000 concurrent requests, the server performs adequately — responses are delivered reliably.
• For >10,000 concurrent requests, the client frequently experiences disconnections or timeouts, indicating instability or resource exhaustion.
• In contrast, bun handles up to 100,000 concurrent requests with consistent performance
• When comparing both at 1,000 concurrent requests bun reports lower latency (67ms bun vs 134ms usockets)
• I tried compiling with IO_URING to see if changing the event loop would improve performance and got the warning below. Despite the warning during compilation, the http_server executable is generated and I tested that similar findings as above.

Question:

Can the uSockets-based http_server example be optimized to match or surpass bun's performance? Such that while sending a simple short text/html response it:

1. Handles more than 100,000 concurrent connections reliably.
2. Matches or exceeds the performance (latency, throughput) of bun — particularly at high concurrency.

URING warning

When I tried to make WITH_IO_URING=1 I get the following warning:

rm -f *.o
cc -flto -DLIBUS_NO_SSL -DLIBUS_USE_IO_URING -std=c11 -Isrc -O3 -c src/*.c src/eventing/*.c src/crypto/*.c src/io_uring/*.c
src/io_uring/io_loop.c: In function ‘us_create_loop’:
src/io_uring/io_loop.c:343:20: warning: passing argument 1 of ‘posix_memalign’ from incompatible pointer type [-Wincompatible-pointer-types]
  343 |     posix_memalign(&reg.ring_addr, 1024 * 4, sizeof(struct io_uring_buf) * 4096);
      |                    ^~~~~~~~~~~~~~
      |                    |
      |                    __u64 * {aka long long unsigned int *}
In file included from src/io_uring/internal.h:5,
                 from src/io_uring/io_loop.c:23:
/usr/include/stdlib.h:718:35: note: expected ‘void **’ but argument is of type ‘__u64 *’ {aka ‘long long unsigned int *’}
  718 | extern int posix_memalign (void **__memptr, size_t __alignment, size_t __size)
      |                            ~~~~~~~^~~~~~~~
src/io_uring/io_loop.c:346:20: warning: assignment to ‘struct io_uring_buf_ring *’ from ‘__u64’ {aka ‘long long unsigned int’} makes pointer from integer without a cast [-Wint-conversion]
  346 |     loop->buf_ring = reg.ring_addr;
      |                    ^
src/io_uring/io_loop.c:343:5: warning: ignoring return value of ‘posix_memalign’ declared with attribute ‘warn_unused_result’ [-Wunused-result]
  343 |     posix_memalign(&reg.ring_addr, 1024 * 4, sizeof(struct io_uring_buf) * 4096);
      |     ^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
lib.exe /out:uSockets.a *.o || ar rvs uSockets.a *.o
/bin/sh: 1: lib.exe: not found
r - bsd.o
r - context.o
r - epoll_kqueue.o
r - gcd.o
r - io_context.o
r - io_loop.o
r - io_socket.o
r - libuv.o
r - loop.o
r - openssl.o
r - quic.o
r - socket.o
r - udp.o
for f in examples/*.c; do cc -O3 -flto -DLIBUS_NO_SSL -DLIBUS_USE_IO_URING -std=c11 -Isrc -o $(basename "$f" ".c") "$f" uSockets.a /usr/lib/liburing.a; done
In function ‘us_socket_timeout’,
    inlined from ‘us_socket_timeout’ at src/io_uring/io_socket.c:52:6,
    inlined from ‘main’ at examples/hammer_test.c:439:5:
src/io_uring/io_socket.c:54:20: warning: writing 1 byte into a region of size 0 [-Wstringop-overflow=]
   54 |         s->timeout = ((unsigned int)s->context->timestamp + ((seconds + 3) >> 2)) % 240;
      |                    ^
In function ‘us_socket_context_listen’,
    inlined from ‘main’ at examples/hammer_test.c:434:21:
src/io_uring/io_context.c:138:43: note: at offset 24 into destination object of size 16 allocated by ‘malloc’
  138 |     struct us_listen_socket_t *listen_s = malloc(sizeof(struct us_listen_socket_t));

[uNetworkingAB]

IO_URING support is experimental and not production ready.

[ImmanuelSamuel]

I edited the question to be more clear with regards to the end goal.

[uNetworkingAB]

Yes uSockets is faster than Bun for the simple reason Bun is a JS environment built using uSockets. Bun is not using IO_URING.

[ImmanuelSamuel]

Does the example http_server run faster than a similar simple text response 'hello world' bun implementation for you? No SSL, just a simple get request. I am having a bunch of disconnections with high concurrent requests and the latency is slower.

[uNetworkingAB]

Yes, a lot. You can use the http_load_test and http_server examples to get a demo of faster perf.

If you're using a JS client to benchmark a C server you need to stop doing that and use a real client instead. http_load_test is one