The Production Problem You’re pushing 500K messages per second between processes and your sys CPU time is through the roof. Your profiler shows mq_send() and mq_receive() dominating the flame graph. Each message is tiny—maybe 64 bytes—but you’re burning 40% CPU just on IPC overhead.

This isn’t a hypothetical. LinkedIn’s Kafka producers hit exactly this wall. Message queue syscalls were killing throughput. They switched to shared memory ring buffers and saw context switches drop from 100K/sec to near-zero. The difference? Every message queue operation is a syscall with user→kernel→user memory copies. Shared memory lets you write directly to memory the other process can read. No syscall after setup, no context switch, no copy.

The performance cliff sneaks up on you. At low rates, message queues work fine—the kernel handles synchronization and you get clean blocking semantics. But scale up and suddenly you’re paying 60-100ns per syscall, plus the cost of copying data twice and context switching when queues block. Shared memory with lock-free algorithms can hit sub-microsecond latencies, but you’re now responsible for synchronization, cache coherency, and cleanup if a process crashes mid-operation.