Tuesday December 06, 2005
alanc @ sun.com
Alan Coopersmith’s blog
Random thoughts of a disorganized mind...
(and though it should be obvious, while Sun pays me to think about things, they disclaim any responsibility for these thoughts, nor do I claim what I say matches in any way what Sun thinks)
Loading...
Tuesday December 06, 2005
Solaris Desktop Summit: Performance Day 1
The Performance portion of the Solaris Desktop Summit got off to a rousing start today with Bryan's talk on the secrets of the DTrace gurus (or at least all the new bits they've added recently and are still working on documenting in the DTrace Guide.
In the afternoon, we broke into working groups, and I joined the team looking at Boot Time. We split our group further, so I worked with gdm maintainer Brian Cameron on trying to see if we could figure out where we could improve the time it takes to start gdm and the X server once the rest of the system is up. We pieced together enough dtrace to determine that about half of the time was spent in gdmlogin, the program that draws the actual login dialog, about a quarter in Xorg itself, and the rest spread across the other processes involved (including the many svcprop calls from the /usr/X11/bin/Xserver script to get the X preferences from the Solaris SMF registry - an area already noted by others as ripe for improvement).
So we started with the gdmlogin process and dug in - it wasn't long before we hit the limit of my DTrace skills, and we called in Bryan to pinch-hit. [To keep things less confusing in the next part, I'll switch to login names, and refer to Bryan as “bmc” and Brian as “yippi.”] He flew through it - finding things that looked strange and digging in, while yippi looked in the source code for explanations. We saved the scripts and DTrace one-liners on yippi's laptop, and will hopefully be able to post more later. There were three main things we found to investigate/try improving:
- bmc noticed a lot of time being spent in the close() system call, which seemed strange. Tracking this down to the stack trace involved, yippi saw it was socket closes in the gdmlogin configuration calls. When gdmlogin needs to get a configuration value from the parent gdm daemon it opens a socket, checks the version, asks for the value it needs, then tears down the socket. Over and over again it does this, instead of simply caching the socket and reusing it for future calls, so yippi is looking into seeing if we can convert it to do so.
- bmc also noticed a lot of stat() system calls, and found two big causes:
- gdm has a list of possible supported locales and stats the locale directories for each ones to see which are installed. We thought of two ways to reduce the cost of these. First, instead of having gdmlogin do this every time, having the parent do it and pass the list to the children could reduce the time on systems with multiple login screens, such as multiseat systems, including Sun Ray. That shouldn't make much difference on single user system though, but the second idea, to change the scan to instead read the parent directory containing all the locales and use the list of it's subdirs instead of a huge list of all possible subdirs could greatly cut down the number of system calls on systems with only a few locales installed.
- gdm was also scanning lots of directories to find the icons for the gtk theme, including stat'ing the same directory over and over again. Another group is looking into gtk performance and will be looking into this more.
We didn't get a chance to make all the changes and test the code — at least not before I left, though yippi was still debugging the first set of changes then and if he didn't finish today, we should be tackling it again tomorrow to see if we can measure improvememnts from the changes and find any more or start digging into Xorg.
P.S. John has written a lot more than I did about the Usability portion of the Summit we held last week, including some of the problems we identifies and the ideas we kicked around to solve them.
[Technorati Tags: Xorg, X11, GNOME, DTrace, Solaris, OpenSolaris.]
Posted at 12:26AM Dec 06, 2005 by Alan Coopersmith in Solaris | Comments[1]
Subscribe!
Posted by Alexander Larsson on December 06, 2005 at 08:12 AM PST #