Meta has reduced its CPU cycles and server demands by up to 20% with Strobelight, a profiling orchestrator powered by eBPF (extended Berkeley Packet Filter) for observability data collection. This solution has driven measurable efficiency improvements across Meta’s infrastructure, resulting in substantial capacity savings and operational benefits.

Tackling the Challenge of Efficient Profiling

Meta faced a critical challenge: gathering detailed performance profiling data across its extensive backend services without disrupting live operations. The company required a solution that could:

• Collect performance data with minimal impact on running services.
• Normalize profiling data for better interpretability.
• Reduce storage burdens caused by excessive data collection.
• Support multiple kernel versions across its infrastructure.

The Solution

To address these challenges, Meta developed Strobelight, a highly efficient profiling system that integrates multiple observability tools, with eBPF as a core component. Strobelight captures critical performance insights, including:

• CPU time usage across execution paths.
• Call stack analysis for various programming languages, including Python, Java, and Erlang.
• Service request latency and off-CPU time tracking.
• AI/GPU profiling and memory usage monitoring.

Unlike traditional profiling tools that require direct instrumentation inside application binaries, Strobelight operates with minimal interference, preserving system performance.

Breakthrough Efficiency Gains

The deployment of eBPF within Strobelight has led to substantial operational improvements:

• Annual capacity savings equivalent to 15,000 servers—achieved from a single one-character code change.
• A 20% drop in CPU cycles, reducing the number of required servers for Meta’s top services by 10–20%.
• Faster debugging and regression detection, preventing performance issues before they impact production.
• Dynamic sampling mechanisms, which optimize profiling rates while avoiding excessive storage demands.

Why eBPF?

eBPF was chosen as the foundation for Strobelight due to its low overhead, flexibility, and powerful kernel-level instrumentation capabilities. Unlike traditional profiling methods, eBPF enables:

• Real-time profiling without modifying application code.
• Wide compatibility across multiple kernel versions, with adaptive fallback mechanisms.
• Efficient data collection with minimal impact on system performance.

Looking Ahead: Expanding eBPF’s Role at Meta

Building on the success of Strobelight, Meta plans to further expand its use of eBPF for AI/ML observability, advanced memory tracking, and deeper efficiency analyses. Additionally, the company is exploring open-sourcing parts of Strobelight, allowing the broader developer community to benefit from its innovations.