Performance Hints (2023)

Summary

Performance Hints Jeff Dean, Sanjay Ghemawat Original version: 2023/07/27, last updated: 2025/12/16 Expand all details Collapse all details Over the years, we (Jeff & Sanjay) have done a fair bit of diving into performance tuning of various pieces of code, and improving the performance of our software has been important from the very earliest days of Google, since it lets us do more for more users. We wrote this document as a way of identifying some general principles and specific techniques that we use when doing this sort of work, and tried to pick illustrative source code changes (change lists, or CLs) that provide examples of the various approaches and techniques. Most of the concrete suggestions below reference C++ types and CLs, but the general principles apply to other languages. The document focuses on general performance tuning in the context of a single binary, and does not cover distributed systems or machine learning (ML) hardware performance tuning (huge areas unto themselves). We hope others will find this useful. Many of the examples in the document have code fragments that demonstrate the techniques (click the little triangles!). Note that some of these code fragments mention various internal Google codebase abstractions. We have included these anyway if we felt like the examples were self-contained enough to be understandable to those unfamiliar with the details of those abstractions. The importance of thinking about performance Knuth is often quoted out of context as saying premature optimization is the root of all evil. The full quote reads: “We should forget about small efficiencies, say about 97% of the time: premature optimization is the root of all evil. Yet we should not pass up our opportunities in that critical 3%.” This document is about that critical 3%, and a more compelling quote, again from Knuth, reads: The improvement in speed from Example 2 to Example 2a is only about 12%, and many people would pronounce that insignificant. The conv...