Global Lock Bench

This repository contains three C benchmarks for measuring spinlock contention and pause instruction behavior on multi-core systems. Each benchmark is designed to highlight different aspects of cache contention and thread synchronization.

Platform Requirements

These benchmarks are designed and tested for x86/x86_64 processors only.

Requirements:

• CPU: x86 or x86_64 architecture
• OS: Linux or similar Unix-like system
• Compiler: GCC with support for inline assembly and atomic builtins
• Features: Uses x86-specific PAUSE instruction and RDTSC for cycle-accurate timing

Note: The code uses x86-specific inline assembly (pause, rdtsc) and has not been tested on other architectures. For cross-platform compatibility, consider this a reference implementation for x86 systems.

Files

• pause_bench.c
  A single-threaded microbenchmark that measures the latency of the pause instruction in a tight loop. Useful for establishing a baseline for pause latency on a given system.

• lock_bench_threaded.c
  A multi-threaded benchmark where all threads contend on a single global spinlock and shared counter. This file demonstrates the effects of lock contention and cache line bouncing as thread count increases.

• lock_bench_padded.c
  Similar to lock_bench_threaded.c, but the spinlock variable is padded and aligned to a cache line to reduce false sharing with adjacent variables. This helps isolate the effect of lock contention from false sharing.

Building

Run the following script to compile all benchmarks:

./build.sh

Requirements:

• GCC compiler with support for inline assembly
• x86/x86_64 target architecture
• POSIX threads (pthread) support

Each executable will be named after its source file:

• pause_bench
• lock_bench_threaded
• lock_bench_padded

Usage

Each benchmark can be run directly from the command line. Most accept optional arguments for thread count and work units. For example:

./lock_bench_threaded 16 10000000
```plaintext

Refer to the source code comments for details on usage and parameters.