在 2024/5/29 07:13, David Sterba 写道:
On Tue, May 28, 2024 at 12:44:19PM +0200, Zaslonko Mikhail wrote:
But I'm still wondering if we do not go 4 pages as buffer, how much
performance penalty would there be?
One of the objective is to prepare for the incoming sector perfect
subpage compression support, thus I'm re-checking the existing
compression code, preparing to change them to be subpage compatible.
If we can simplify the behavior without too large performance penalty,
can we consider just using one single page as buffer?
Based on my earlier estimates, bigger buffer provided up to 60% performance for inflate and up to 30% for
deflate on s390 with dfltcc support.
I don't think giving it away for simplification would be a good idea.
60% and 30% sound like significant gain, I agree this takes precedence
over code simplification. Eventually the s390 optimized case can be
moved to a separate function if the conditions are satisfied so it's not
mixed with the page-by-page code.
Thanks a lot for the numbers, the number is indeed impressive.
I'll definitely leave the larger buffer support untouched.
And thankfully since S390 only supports 4K page size, we won't need to
bother too much for that routine.
Thanks,
Qu
