Re: 1.2.13 v 1.3.81 open() speed results

Leonard N. Zubkoff (lnz@dandelion.com)
Thu, 18 Apr 1996 13:19:28 -0700


Date: Tue, 16 Apr 1996 15:48:02 +0100 (BST)
From: Dirk-Jan Koopman <djk@chbi.co.uk>

On Tue, 16 Apr 1996, Leonard N. Zubkoff wrote:

> If someone would kindly point me at the kernel profiling app, I will
> recompile and profile the kernel.
>
> Don't bother. I'm 99.99% certain you'll find that get_empty_inode is the
> loser. I've already reported this and run some profiles for Linus...
>

Good, I had a cursory look at the fileopen code; is there any profit on
the old idea of a 'half open' chain. This was a George 3 idea which
effectively hid 'closed' files on double headed chain of finite length.
This provided a secondary cache of highly reusable files such directories.

The idea of course is to search the 'half open' chain for a filename first.

1.3.91 now includes a fix from Linus for the performance degradation I
encountered. You might want to re-run your test now and see if the problem is
still present. 1.3.91 now beats 1.2.13 for me on the test I was running.

Leonard