|
|
Subscribe / Log in / New account

Real customer

Real customer

Posted Jun 5, 2024 21:43 UTC (Wed) by willy (subscriber, #9762)
Parent article: Measuring and improving buffered I/O

Not my customer, and not a large customer. But 64-byte reads are a real workload and this customer notices the performance of the atomic inc/dec on the folio refcount.


to post comments

Real customer

Posted Jun 6, 2024 22:05 UTC (Thu) by Heretic_Blacksheep (guest, #169992) [Link] (1 responses)

I'm curious if this would be a natural result of 64k page sizes common on ARM & Power?

Real customer

Posted Jun 6, 2024 22:10 UTC (Thu) by willy (subscriber, #9762) [Link]

The customer is using x86, so 4KiB page size. The problem would be worse with a 64KiB page size as there would be more contention.


Copyright © 2025, Eklektix, Inc.
Comments and public postings are copyrighted by their creators.
Linux is a registered trademark of Linus Torvalds