Re: [DISCUSSION] FILE* slowness

[apache2]

On Sun, Jul 17, 2022 at 02:10:04PM +0430, Mohammad-Reza Nabipoor wrote:
> Thanks for the work.
> Can you put your ZFS file in the mem IOS and run the benchmark?
> Maybe the main problem is the overhead of IOS subsystem; we need to
> know how much it is.
> 
> If the difference between mem IOS and file IOS is not that much,
> then the problem is not the file IOS!
> 

I don't know how to do that :-(

The main overhead seems to come from copying tiny amounts of bytes, somewhere.
That's the case for both the current FILE* implementation and the proposed 
patch.
This patch only tries to make the file reading faster; after applying it it 
still takes about 1 minute to parse a 45 kilobyte file.

I can try to get some more details caller information for the AVX functions 
featured at the top of the perf reports (around 12-15%), but I think that is a 
separate issue from the I/O layer being slow.