It is not ok that we speed weeks, even months, trying to solve why software is slow. Companies waste money on compute costs, users are unhappy with latency, and product evaluations run out of investigation time. It should not take more than a week to identify the root cause or causes for a performance issue, such that any performance issue reported on a Monday should be solved by Friday, or sooner. The kernel superpowers we have been building are essential for this dream, and allow us to explore performance analysis methodologies to achieve this that were previously a fantasy.

This talk explores the dream of “fast by Friday,” and shows how kernel technologies like eBPF, and performance methodologies, can get us there. The end goal is not more tools and metrics or having everyone learn eBPF bytecode. It’s about efficient computing, and solving inefficiencies as quickly as possible. It’s about saving cycles and carbon.

To be fast by Friday requires observability tools to work on Monday, and right now for many Linux environments that means /proc based tools and Ftrace, sometimes perf, and rarely the eBPF tracing tools: bcc and bpftrace. This and other current and future technical challenges will be discussed, including eBPF stack walking, runtime behavior and uprobes, compiler optimization defaults, OS default packages, and non-CPU targets (GPUs, accelerators).

Brendan GREGG

свернуть