This article explains why a TCP connection's Send‑Q can exceed the user‑set SO_SNDBUF value, detailing the kernel's double‑buffer trick, sk_wmem_queued accounting, tcp_sendmsg behavior, GSO influence, and possible ways to limit the buffer growth.
A Linux client sets SO_SNDBUF to 4096 bytes, sends 1 KB packets while the server never reads, yet the TCP Send‑Q grows to 14480 bytes because the kernel doubles the buffer size, adds overhead, and expands memory via GSO and sk_wmem scheduling.
A Linux client sets SO_SNDBUF to 4096 bytes and sends 1 KB packets while the server never receives them, yet the TCP Send‑Q grows to 14480 bytes, prompting an in‑depth analysis of kernel buffer accounting, double‑sized send buffers, and tcp_sendmsg behavior.
This article explains why the TCP Send‑Q can grow beyond the user‑set SO_SNDBUF value by examining the kernel’s double‑buffer accounting, sk_wmem_queued handling, and the interaction of tcp_sendmsg with GSO, supported by code excerpts and diagrams.
This article explains UDP packet structure, why UDP lacks a send buffer, how fragmentation and buffer overflows cause packet loss, and presents practical solutions—including socket buffer tuning, delayed sending, retransmission strategies, redundancy, and RUDP—to achieve reliable communication over an inherently unreliable protocol.
This article explains the Linux network I/O architecture, detailing how TCP sends and receives data, the role of socket buffers, SKB structures, QDisc and ring buffers, and compares various zero‑copy techniques such as read/write, mmap/write, sendfile, splice, tee, DPDK and DMA‑gather approaches.
A 100 M dedicated line with 20 ms latency caused a 22 MB SQL query to take over 25 seconds, prompting an in‑depth analysis of how socket send/receive buffers, sysctl rmem/wmem settings, and TCP windows interact and affect network performance.
