Post

I occasionally wonder how much of the polled network/storage performance gains being talked about are in reality mainly about the effect of CPU idling.

20/2/25, 4:46 pm0 reposts1 like

Andres Freund

@anarazel.de

20 Feb 2025

E.g. news.ycombinator.com/item?id=4311... shows that disabling cpu idling, on server HW, can significantly reduce the time in which a request is processed.

Andres Freund

@anarazel.de

20 Feb 2025

If I instead make the server busy-poll for network activity (by just loopping around non-blocking recv), while waiting for the next query, I get even better numbers than shown there. Even though it's still interrupt based network processing.