TechPrufer
Sign inGet started

Blogs›Cloudflare›Branch Prediction Performance

Branch Prediction Performance

Branch Prediction Performance

1
posts
2021

Cloudflare's understanding of branch prediction performance has evolved, moving beyond the general rule that predictable branches have near-zero cost. Experiments have revealed that the cost of branches, even unconditional ones, can increase significantly with their number and density. This has led to a deeper investigation into the Branch Target Buffer (BTB) and its capacity, demonstrating that performance degrades when the number of branches exceeds certain thresholds (e.g., 4096 branches). This understanding informs code optimization strategies to avoid performance penalties associated with excessive branching.

2021

Branch predictor: How many "if"s are too many? Including x86 and M1 benchmarks!

5/6/2021

This post details experiments investigating the performance impact of numerous branch instructions on modern CPUs, specifically focusing on the Branch Target Buffer (BTB). It demonstrates that the cost of a branch instruction increases as the number of branches grows, especially when exceeding the BTB's capacity. Benchmarks on x86 (AMD EPYC) and M1 architectures show performance degradation beyond approximately 4096 branches, with unpredicted branches costing significantly more cycles. The post also explores the impact of code density and block size on branch prediction performance.