Re: [PATCH] Invalidate_inodes can be very slow

[Nikita Danilov <Nikita@Namesys.COM>]

Kirill Korotaev writes:
 > > At some point in the past, I wrote:
 > > >> Sorry if I was unclear, I had in mind SMP performance testing of mount
 > > >> and unmount -heavy workloads, like uni setups with many automounted
 > > >> fs's, not stability testing per se.
 > >
 > > On Mon, Oct 13, 2003 at 04:02:20PM +0400, Kirill Korotaev wrote:
 > > > Oh, sorry for misunderstanding.
 > > > In our internal testcase on 8-CPU 8Gb RAM machine with 4gb split kernel
 > > > w/o this patch mount/umount test longs in many-many (>10) times longer.
 > > > Moreover, during the test machine is very slow (due to lock_kernel)
 > > > and typing simple commands takes up to 30 seconds or so.
 > > > I think such a long hangs are due to number of umounts executed
 > > > subsequently. But ofcourse it's not numbers, just for you to know where
 > > > the patch comes from :)
 > >
 > > Is this testcase available and/or trivial? Actually, even if it's trivial
 > > it might just save us the pain of writing the scripts ourselves.
 > no, testcase is not available :( And it uses functionality
 > not available in mainstream kernel. But the problem can be hit with
 > very simple script instead:
 > 
 > 1. mount N filesystems.
 > 2. work on them, so that inode cache grows to its maximum
 > possible size (it was 1,000,000 of inodes in our case).
 > 3. umount these filesystems.
 > 
 > During operation #3 node is very slow and it is quite noticable
 > on ssh console when typing commands.

This can be due to a number of reasons (worst case behavior of
shrink_dcache_parent() for example). What /proc/profile shows?

 > 
 > Kirill

Nikita.

 >