Simulating 20,000 Concurrent WebSocket Connections on Linux Kernels

Stress testing real-time WebSocket connections introduces unique OS-level challenges compared to standard HTTP requests. While HTTP connections open, transfer data, and close, WebSockets remain open indefinitely, rapidly exhausting Linux file descriptors and socket memory.

The File Descriptor Limit (nofile)

By default, Linux limits a single process to 1,024 open file descriptors. Because every TCP socket connection consumes one file descriptor, a load generator process will crash almost immediately (java.net.SocketException: Too many open files).

Kernel Tuning Parameters

To allow a single worker instance to handle 20,000+ open connections, adjust /etc/security/limits.conf and sysctl.conf:

# Raise maximum open files for worker process
* soft nofile 65535
* hard nofile 65535

# Expand local port range for ephemeral TCP outbound sockets
sysctl -w net.ipv4.ip_local_port_range="1024 65535"

# Enable socket reuse to prevent TIME_WAIT saturation
sysctl -w net.ipv4.tcp_tw_reuse=1

Socket Memory Optimization (rmem / wmem)
Every open TCP socket allocates read and write buffers in RAM. At 20,000 idle connections, default memory allocations consume several gigabytes of kernel memory.

Default: 128 KB per socket * 20,000 = ~2.5 GB RAM (Kernel Space)
Tuned:    16 KB per socket * 20,000 = ~320 MB RAM (Kernel Space)
Lowering net.ipv4.tcp_rmem and net.ipv4.tcp_wmem minimum values allows long-lived idle WebSocket connections to consume minimal OS memory without dropping incoming frames.