Memory illusions: The way memory ruins your benchmark / Shoham Baris

We will look at a real-life scenario in which, surprisingly, a benchmark tested on a bigger dataset was quicker than on a smaller dataset. The underlying reason is found by exploring the trend of "ballast" from "Twitch", which solves an even intensified problem when benchmarking. To thoroughly understand that topic, a small introduction is required going over the Go memory allocation on top of memory pages, and the development of the Go garbage collection across the latest versions. This issue is not solely due to the memory allocation but is also affected by "Memory alignment" and "Cache hit: tactics, which can dramatically alter the benchmark results. Finally, we'll conclude with some advice on how to benchmark avoiding those caveats. Shoham Baris Software engineer A 6-year software engineer, currently employed at Imubit, embedding AI models into the "backend". I am passionate about the internals of go and memory management in general.