Re: [PATCH 1/3] sysfs directory scaling: rbtree for dirent name lookups

[ebiederm@xmission.com (Eric W. Biederman)]

Greg KH <greg@kroah.com> writes:

> On Tue, Nov 03, 2009 at 07:14:33AM +0100, Eric Dumazet wrote:
>> Greg KH a ?crit :
>> > On Sun, Nov 01, 2009 at 11:31:30AM -0500, Benjamin LaHaise wrote:
>> >> Use an rbtree in sysfs_dirent to speed up file lookup times
>> >>
>> >> Systems with large numbers (tens of thousands and more) of network 
>> >> interfaces stress the sysfs code in ways that make the linear search for 
>> >> a name match take far too long.  Avoid this by using an rbtree.
>> > 
>> > What kind of speedups are you seeing here?  And do these changes cause a
>> > memory increase due to the structure changes which outweigh the
>> > speedups?
>> > 
>> > What kind of test are you doing to reproduce this?
>> > 
>> 
>> Its curious because in my tests the biggest problems come from
>> kernel/sysctl.c (__register_sysctl_paths) consuming 80% of cpu
>> in following attempt to create 20.000 devices
>> 
>> (disable hotplug before trying this, and ipv6 too !)
>> modprobe dummy numdummies=20000
>> 
>> I believe we should address __register_sysctl_paths() scalability
>> problems too.
>
> But registering 20000 devices is a far different problem from using
> those 20000 devices :)
>
> I think the "use the device" path should be the one we care the most
> about fixing up, as that is much more common than the register path for
> all users.

Definitely.  Of the three proc sysctl and sysfs.  sysctl tends to have
the worst costs across the board.  They are all rarely used so a lot
of what gets hit when scaling are rare path events that even the most
horrible code works fine on small systems.

Usually slow registration times indicate an O(N^2) or worse data
structure for filename lookup.

Eric