Investigatoin of Linux Performance/Responsiveness Regressions
Hey, just thought I'd enumerate a bunch of noted slowdowns in Linux I and other people have discovered. I'm relying on some of the numbers from Ubuntu (mainly due to Phoronix's popular recent articles comparing performance), but I'm sure they apply to other distros just as well.
Having talked to some devs and noting that most of the performance loss is introduced between 7.10 and 8.04, I'm wondering if it's due to the CFS task scheduler introduced in 2.6.23 (kernel between 7.10 and 8.04).
Other word on the street (confirmed by my own experiments) is that heavy disk I/O stops all cpu activity on currently running tasks. I wonder if it's due to the new I/O scheduler (elevator option at boot). I've heard that using deadline instead of CFQ allows responsive multitasking again.
There is also a known regression with the intel open-source video driver that is resulting in lower and lower performance on 8.04, 8.10, and 9.04 alpha.
The ultimate goal is to question why we are seeing these serious and noticeable performance regressions. Devs don't usually take notice because we don't have enough clout or hard metrics (just lots of anecdotal evidence). Let's band together to solve this.
Yeah, I've noticed Firefox become unresponsive and my panel-mounted CPU graph shows 100% wait time for several seconds. I have a 4-core processor, but this was after upgrading to Ubuntu 8.10, before I found out the realtime kernel they use disables SMP. After changing back to a normal SMP kernel I still believe I've seen the problem, just less pronounced because there's other processors available - I'll keep an eye out for hard symptoms...
I'm using an ATI card with the Ubuntu-packaged fglrx 8.10 driver, and it's just happened again. Started up firefox, loaded some pages, and a minute or so later firefox freezes and the system hits one processor's worth of wait time, for somewhere between 5 and 10 seconds. The rest of the system was more responsive though, presumably since the other CPUs were free.
Originally Posted by kraftman
I guess the "nopat" option in GRUB boot line fixes this problem.
Originally Posted by grantek