| Syscall | Cost | Description | When it appears |
|---|---|---|---|
read() | ~100–500ns | Read from fd (returns if data ready, blocks otherwise) | Every network recv, file read |
write() | ~100–500ns | Write to fd | Every network send, log write |
accept() | ~500ns–1μs | Accept new connection from listen socket | New connection per request (no pool) |
connect() | ~500ns + RTT | Initiate TCP connection | New connection creation |
epoll_wait() | ~500ns | Wait for events on registered fds | Event loop at rest |
epoll_ctl() | ~500ns | Add/modify/remove fd from interest list | On every new connection (async server) |
socket() | ~1μs | Create new socket fd | Connection creation |
close() | ~500ns | Close fd + optional TCP teardown | Connection close |
fork() | ~1–5ms | Create new process | PostgreSQL: per connection |
mmap() | ~1μs | Map file/anonymous memory into address space | Buffer pool creation |
brk()/sbrk() | ~500ns | Extend heap | malloc() for large allocations |
futex() | ~50–200ns | Fast user-space mutex (contended case) | Any mutex contention |
Post-Spectre/Meltdown (2018), syscalls are more expensive due to KPTI (Kernel Page Table Isolation):
bash
in this section