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

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

Benjamin LaHaise <bcrl@lhnet.ca> writes:

> On Tue, Nov 03, 2009 at 01:32:33PM -0800, Eric W. Biederman wrote:
>> Are your numbers from your application and are they real world?
>> In which case they are interesting, but it would be good if
>> we could also have microbenchmark numbers that just measure
>> the sysfs costs.   If nothing else I am seeing a big startup
>> overhead that isn't being subtracted out that makes it hard
>> to see the real costs here.
>
> They're application based, so there's a bunch of other overhead included 
> that won't show up on a microbenchmark.  Each interface requires a round 
> trip between 2 L2TP daemons, so there are lots of syscalls and other cache 
> polluting effects that won't show up on a microbenchmark.  One of the L2TP 
> daemons is configured not to instantiate any kernel state -- running in 
> this mode, it has very little overhead.
>
> The other thing to note is that the costs posted are how long it takes to 
> add an additional 5,000 interfaces in the given range, not the total time 
> to add say 35,000 interfaces (I didn't feel like waiting that long).

Ok.  That makes a lot more sense.  The times you posted ideally would be flat
but they go up from 12s to 60s.

Eric