Principles of Mechanical Sympathy
3 days ago
- #concurrency
- #hardware-optimization
- #software-performance
- Mechanical sympathy involves designing software to work efficiently with hardware, using principles like predictable memory access, cache line awareness, single-writer principle, and natural batching.
- CPUs organize memory hierarchically (registers, L1/L2/L3 caches, RAM); predictable, sequential access outperforms random access due to CPU prefetching strategies.
- Cache lines (e.g., 64-byte chunks) can cause false sharing when multiple CPUs write to different variables in the same line, degrading performance; padding cache lines can mitigate this.
- The single-writer principle assigns all writes to a specific resource (e.g., a variable or socket) to one thread, reducing mutex contention and enabling optimizations like batching.
- Natural batching improves latency by having a single writer start batches as soon as data is available and complete them when the queue is empty or full, avoiding fixed timeouts.
- These principles scale from individual applications to distributed systems, but observability (SLIs, SLOs, SLAs) should be prioritized before optimization.