Improving Server Concurrency and Performance: Methods and Strategies

Server concurrency refers to the number of requests a server can handle per unit time; higher concurrency means higher capacity.

Two common metrics are throughput (requests per second) and stress testing, which evaluates performance under varying numbers of simultaneous users, total request count, request resources, and waiting times.

Improving CPU concurrency involves using multiple processes and threads, leveraging DMA to offload I/O, and reducing context switches by limiting process count or binding processes to CPUs.

Minimizing lock contention, using lock‑free programming with atomic operations, and disabling unnecessary logging can further reduce waiting time.

Persistent (keep‑alive) connections reduce the overhead of establishing TCP connections, especially for small, frequent requests.

Optimizing I/O models—such as employing DMA, asynchronous I/O, epoll, sendfile, memory‑mapped files ( mmap ), and direct I/O ( O_DIRECT )—allows the CPU to overlap computation with slower I/O operations.

Different concurrency strategies include one process per connection (prefork), one thread per connection (worker), and asynchronous models where a single process or thread handles many connections using I/O multiplexing (epoll, edge‑triggered, etc.).

Finally, scaling hardware (adding CPU cores, faster network adapters, RAID storage) provides a straightforward way to increase capacity.

The content is based on the article "构建高性能Web站点" and the author's original blog post.