Backend Performance Study

Built a microbenchmark suite to study how cache locality, false sharing, and lock contention affect service latency and throughput under concurrency, and to analyze root causes of latency spikes in a storage data plane.

What I did

• Designed benchmarks to isolate cache locality patterns, false sharing between threads, and lock contention under different workload profiles
• Used perf to measure CPU cycles, cache misses, and scheduler effects across single-threaded and multi-threaded code paths
• Used valgrind tooling to profile memory behavior and identify bottlenecks
• Applied optimizations such as improving data locality, cache-line padding, and memory alignment

What I learned

The project made systems tradeoffs more concrete — especially around concurrency and memory-access patterns. Understanding how the CPU cache hierarchy affects real-world service performance changed how I think about backend data structure design.