Question 1

What is the C10K problem and how was it solved?

Accepted Answer

The challenge of handling 10,000 concurrent network connections efficiently — solved with event-driven I/O (epoll/kqueue) instead of one OS thread per connection. Thread-per-connection: 10K threads × ~8MB stack = 80GB RAM + massive context switching. Solution: single-threaded event loop + non-blocking I/O via epoll. nginx, Node.js, and Redis are all based on this model. Coined by Dan Kegel in 1999.

Question 2

How does epoll differ from select for I/O multiplexing?

Accepted Answer

epoll provides O(1) event notification regardless of the number of watched file descriptors; select has O(n) scan overhead and a fixed FD_SETSIZE limit (~1024). select: the kernel scans all n FDs every call — O(n). epoll: register interest once with epoll_ctl(); epoll_wait() returns only the FDs that are ready — O(events). For 10K connections with 100 active, epoll wakes up with 100 events; select scans all 10K.

Question 3

What is NUMA (Non-Uniform Memory Access) and how does it affect performance?

Accepted Answer

A memory architecture where each CPU socket has its own local memory bank (fast access) and can access other sockets' memory (slower) — locality is critical for performance. On a 2-socket server: CPU 0 accessing its own RAM takes ~80ns; accessing CPU 1's RAM takes ~120ns. Performance-critical software (databases, JVMs) pins threads and memory to the same NUMA node. Use numactl to control NUMA allocation.

Question 4

What is a memory-mapped file (mmap) and when is it preferable to read()/write()?

Accepted Answer

Maps a file directly into the process's virtual address space — file I/O becomes memory accesses; useful for large files, random access, and zero-copy IPC between processes. read(): user buffer ← kernel buffer ← disk (two copies). mmap(): file is the buffer — one copy, and only pages actually accessed are loaded. Databases (SQLite, PostgreSQL) use mmap for reading data files. Also enables multiple processes to share memory via MAP_SHARED.

Question 5

What is CPU affinity and when would you set it?

Accepted Answer

Binding a process or thread to specific CPU cores — improves cache locality and reduces context switching overhead for latency-sensitive workloads. Use taskset or pthread_setaffinity_np(). Useful for: latency-critical apps (trading systems), isolating noisy neighbors in the cloud, reducing NUMA cross-socket traffic. Combined with isolcpus kernel param to dedicate cores exclusively to your process.

Operating Systems — Senior Dev Interview Questions

More Senior Dev Topics

Operating Systems at Other Levels