[Users] [Bug 2584] GUI redraws itself slowly - solved for me

Michael codejodler at gmx.ch
Tue Jan 24 03:35:38 CET 2012


I have to say this with care, because i may be still on err and did not much testing yet.

But Pauls' and Slavkos' comments made me experiment with filesystem I/O and i discovered that the 3.x kernel (from Debian) seems to enable the option 'barrier=1' per default now, not only for ext4 but also for ext3, which definitely was not the case in the past. After reading about what this all is, i decided i don't need it (except for some mission critical archives) and i added some clever speed enhancing mount options to most partitions of /etc/fstab, namely "barrier=0,data=writeback,nodelalloc".

And guess what, no delays anymore, not by workspace switching, nor message display, and not even when closing windows.

This is not the first thing i'm doing to speed up file systems. I'm used to apply options like noatime, nodiratime for anything but the root-fs (/), and /tmp and /var/run and /var/lock are shifted to RAM (tmpfs) since long. Then, there is a complex partitioning layout with /var and /var/cache and about anthing on some dozen separate partitions, on 4 SATA harddrives, for parallel access and also to reduce fragmenting. Some filesystems have dedicated design (like large_file) and i even considered the partitions position on the physical disk. All in all my daily system is heavily speed optimized and i am not used to delays at all. I get nervous when things don't happen within 1/3 second ;) 

That said, i was really totally surprised things like redrawing windows or closing applications need that much file system access. But the fact that KDE does it differently and that gsd makes it worse proves (to me) that GTK somehow is doing exactly this. The change in the FS driver defaults only made that apparent. Why on earth would redrawing a window require file system ascess ? I'm totally stunned. 

But it is not the first time i saw change of one factor casually modify some other factor too (for example, i re-ran update-grup to rebuild the initrd) and as i said, anything preliminary, but i just wanted to immediately share this discovery. Perhaps someone else might try to confirm.



More information about the Users mailing list