Question

Please, Linux kernel hackers, what is a reasonable buffer size for write(2) syscall to sockets or files, performance-wise? It’s clear that it’s some pagesize multiple, but which one? Does it matter? What is “too small” and “too big”?

Answer 1

depends on how big your delay to the peer is, let’s say you have a 100MBps connection, and a delay of 50ms, then you can calculate

100MBps * 0.050 sec / 8 = 0.625MB = 625KB

but the default window size in Linux 2.6 is around 110KB, which will limit your throught out to around 2.2MBps (110KB / 0.050)
so, to fix that you can you setsockopt

int ret, sock, buf_size;
sock = socket(AF_INET, SOCK_STREAM, 0);
buf_size = 625*1024;
ret = setsockopt(sock, SOL_SOCKET, SO_SNDBUF, (char *)&buf_size, sizeof(buf_size));
ret = setsockopt(sock, SOL_SOCKET, SO_RCVBUF, (char *)&buf_size, sizeof(buf_size));