Re: kern/55268: tmpfs is slow

[Alexander Nasonov <alnsn%yandex.ru@localhost>]

Joerg Sonnenberger wrote:
> The following reply was made to PR kern/55268; it has been noted by GNATS.
> 
> From: Joerg Sonnenberger <joerg%bec.de@localhost>
> To: gnats-bugs%netbsd.org@localhost
> Cc: 
> Subject: Re: kern/55268: tmpfs is slow
> Date: Fri, 15 May 2020 18:29:47 +0200
> 
>  On Fri, May 15, 2020 at 03:55:00PM +0000, Andreas Gustafsson wrote:
>  > Doing a "build.sh -j 24 release" on a tmpfs takes more real time than
>  > on an ffs mounted with "-o async", and uses much more system time.
>  
>  Careful with DIAGNOSTIC as it increases vnode lock contention quite a
>  bit. Don't have the performance numbers for the bulk build anymore, but
>  it is quite visible. Otherwise, this greatly suggests some form of lock
>  contention or another, so lockstat output would be the most useful thing
>  (e.g. lockstat -b 999999 build.sh ...). This doesn't reflect my
>  experience at the very least, but lock contention can be somewhat
>  catastrophic and small changes can have dramatic impact. That said, in
>  the current form the bug report is not really useful.

Andreas stated that DIAGNOSTIC and acpicpu are DISABLED.

This is very similar to my setup but I run on AMD EPYC. When I was
measuring build performance on tmpfs, I often ran lockstat and I
sent some of those reports to ad@ for analysis. He may have some
ideas what's going on there.

-- 
Alex