二、io_uring 版本异步 send/recv 压测io_uring 核心把send/recv提交到SQ内核异步执行批量收割CQE。减少系统调用次数高性能网络首选。网络io_uring有两种模式IORING_OP_SEND/IORING_OP_RECV原生异步send/recv最简单下面示例io_uring 多连接模式 acceptIORING_OP_ACCEPT⚠️ 编译需要链接 liburingapt install liburing-dev/yum install liburing-develio_uring_server.c#include stdio.h #include stdlib.h #include string.h #include unistd.h #include sys/socket.h #include netinet/in.h #include liburing.h #include time.h #define PORT 8888 #define BUF_SIZE (120) #define QUEUE_DEPTH 16 static int set_nonblock(int fd) { int flags fcntl(fd, F_GETFL, 0); return fcntl(fd, F_SETFL, flags | O_NONBLOCK); } int main() { int listen_fd socket(AF_INET, SOCK_STREAM, 0); set_nonblock(listen_fd); struct sockaddr_in addr {0}; addr.sin_family AF_INET; addr.sin_addr.s_addr INADDR_ANY; addr.sin_port htons(PORT); bind(listen_fd, (struct sockaddr*)addr, sizeof(addr)); listen(listen_fd, 128); struct io_uring ring; io_uring_queue_init(QUEUE_DEPTH, ring, 0); char *buf malloc(BUF_SIZE); struct io_uring_sqe *sqe; struct io_uring_cqe *cqe; int conn_fd -1; long long total_recv 0; clock_t start 0; // submit accept sqe io_uring_get_sqe(ring); io_uring_prep_accept(sqe, listen_fd, NULL, NULL, 0); io_uring_submit(ring); while (1) { io_uring_wait_cqe(ring, cqe); if (cqe-res 0) { perror(cqe err); io_uring_cqe_seen(ring, cqe); continue; } if (conn_fd 0) { // accept完成 conn_fd cqe-res; printf(io_uring client connected fd%d\n, conn_fd); start clock(); io_uring_cqe_seen(ring, cqe); // 提交recv sqe io_uring_get_sqe(ring); io_uring_prep_recv(sqe, conn_fd, buf, BUF_SIZE, 0); io_uring_submit(ring); } else { // recv完成 int n cqe-res; io_uring_cqe_seen(ring, cqe); if (n 0) { clock_t end clock(); double sec (double)(end - start)/CLOCKS_PER_SEC; printf(io_uring server recv total: %lld bytes, bw: %.2f MB/s\n, total_recv, total_recv / sec / (120)); close(conn_fd); free(buf); io_uring_queue_exit(ring); close(listen_fd); return 0; } total_recv n; // 继续提交下一次recv sqe io_uring_get_sqe(ring); io_uring_prep_recv(sqe, conn_fd, buf, BUF_SIZE, 0); io_uring_submit(ring); } } }io_uring_client.c#include stdio.h #include stdlib.h #include string.h #include unistd.h #include sys/socket.h #include netinet/in.h #include arpa/inet.h #include liburing.h #include time.h #define PORT 8888 #define BUF_SIZE (120) #define SEND_TOTAL (10LL 30) #define QUEUE_DEPTH 16 int main() { int sock socket(AF_INET, SOCK_STREAM, 0); struct sockaddr_in serv_addr {0}; serv_addr.sin_family AF_INET; serv_addr.sin_port htons(PORT); inet_pton(AF_INET, 127.0.0.1, serv_addr.sin_addr); connect(sock, (struct sockaddr*)serv_addr, sizeof(serv_addr)); struct io_uring ring; io_uring_queue_init(QUEUE_DEPTH, ring, 0); char *buf malloc(BUF_SIZE); memset(buf, A, BUF_SIZE); long long remain SEND_TOTAL; clock_t start clock(); struct io_uring_sqe *sqe; struct io_uring_cqe *cqe; // 提交第一次send sqe io_uring_get_sqe(ring); int send_len remain BUF_SIZE ? BUF_SIZE : remain; io_uring_prep_send(sqe, sock, buf, send_len, MSG_NOSIGNAL); io_uring_submit(ring); while (remain 0) { io_uring_wait_cqe(ring, cqe); if (cqe-res 0) { perror(send cqe err); exit(1); } int n cqe-res; io_uring_cqe_seen(ring, cqe); remain - n; if (remain 0) break; // 继续提交send sqe io_uring_get_sqe(ring); send_len remain BUF_SIZE ? BUF_SIZE : remain; io_uring_prep_send(sqe, sock, buf, send_len, MSG_NOSIGNAL); io_uring_submit(ring); } clock_t end clock(); double sec (double)(end - start)/CLOCKS_PER_SEC; printf(io_uring client send total: %lld bytes, bw: %.2f MB/s\n, SEND_TOTAL, SEND_TOTAL / sec / (120)); close(sock); free(buf); io_uring_queue_exit(ring); return 0; }编译运行gcc io_uring_server.c -o io_uring_server -O2 -luring gcc io_uring_client.c -o io_uring_client -O2 -luring # 窗口1 ./io_uring_server # 窗口2 ./io_uring_clientio_uring关键点QUEUE_DEPTHSQ/CQ队列深度对性能影响很大小包场景调大可以批量提交io_uring_submit()可以攒一批SQE一次性提交减少syscall高级使用IORING_OP_SEND_ZC零拷贝send内核5.18进一步减少内存拷贝。三、压测指标采集 perf 分析必做# 看CPUsys占比是重点 perf stat ./io_uring_server # 火焰图看调用栈 perf record -g ./io_uring_server perf script | ./FlameGraph/stackcollapse-perf.pl | ./FlameGraph/flamegraph.pl io_uring.svg采集指标吞吐 MB/sserver recv字节为准系统CPU%epoll小包会sys很高io_uring因为批量提交sys开销更低平均延迟改成ping-pong小包测试内存占用、socket buffer溢出ss -ti看tcp队列内核调优网络压测sysctl -w net.core.rmem_max33554432 sysctl -w net.core.wmem_max33554432 sysctl -w net.ipv4.tcp_rmem4096 87380 33554432 sysctl -w net.ipv4.tcp_wmem4096 65536 33554432四、epoll vs io_uring 对比总结特性epollio_uring模型等待fd就绪就绪后调用send/recv异步提交send/recv内核完成后返回CQEsyscall开销每次收发都syscall小包差可批量提交SQEsyscall更少内存拷贝用户→内核拷贝支持send_zc零拷贝适用场景大量长连接连接数优先高吞吐、低sys开销高性能IO参考链接 0voice · GitHub