Tool: macOS-memory-benchmark (Open Source auf meinem GitHub) Daten: Gemessene Direktzugriffslatenz auf einem Apple M4-Chip. Methodik: Das Tool führt Speicherzugriffsmuster in 1024-KB-Schritten aus, um die Latenzschritte der L-Caches und des System Level Cache (SLC) abzubilden. Einblicke: Sie können sehen, wie der SLC-Übergang bei 16 MB beginnt und die DRAM-Latenz bei etwa 40 MB vollständig erreicht.
Von qettyz
3 Kommentare
Source: ran macOS-memory-benchmark tool on my Mac mini M4 24GB (base). There is python script along with sh-script in my GitHub repository what was used to create image from json-files provided by macOS-memory-benchmark -tool.
My advice would be: don’t settle for the default font size, line width, etc., in a plot like this.
You will convey an understanding of your data much better with much larger font size for title and axis labels and axis values, with much thicker lines, with larger symbols along each line, with a more pronounced and better placed legend, and with fewer, larger labels along each axis.
Check a graph that you like and notice the differences in the style that they have chosen to employ and the one that you have used.
e.g., [https://images.squarespace-cdn.com/content/v1/5b872f96aa49a1a1da364999/0298f778-d22c-45f8-acfd-57ecfb55afb7/reaction_rate_graph.png?format=1000w](https://images.squarespace-cdn.com/content/v1/5b872f96aa49a1a1da364999/0298f778-d22c-45f8-acfd-57ecfb55afb7/reaction_rate_graph.png?format=1000w)
from this random blog that I just found: [https://about.dataclassroom.com/blog/multiple-line-graphs](https://about.dataclassroom.com/blog/multiple-line-graphs)
You’re not matching the read size to cache lines so you’re not measuring cache latency at all. Some reads will end up straddling cache lines unpredictably, so instead of a nice plot of latency, you get this weird slope thing.