Single Priority Array CPU Scheduler

Hi Peter,

I just started to try out your SPA scheduler patch and found that it is noticeably sluggish when resizing a mozilla window on the desktop. I have a profile of 2.6.7-rc3-spa and 2.6.7-rc2-mm2 and put them up at:
http://zeke.yi.org/linux/spa/ . There is also vmstat output there but it doesn't look too helpful to me.

The test was basic and went like this:

x86, K7, UP, gnome desktop with mozilla (with a bunch of tabs) and a few rxvts. cmdline= elevator=cfq profile=1

readprofile -r

grab a corner of my mozilla window and continually move it around for several seconds

readprofile -v -m /boot/System.map-2.6.7-rc3|sort -rn +2|head -n30

do the same while dumping vmstat 1 to a file.

The kernel with your patch had a much harder time keeping up with the window resizing. Moving the entire window did not seem too bad or not too noticeable. I tried a similar test while running a kernel compile (make -j3) and it made the window resizing _really_ slow to respond.

Quoted 8 lines · Shane Shrybman

Hi Peter,

I just started to try out your SPA scheduler patch and found that it is noticeably sluggish when resizing a mozilla window on the desktop. I have a profile of 2.6.7-rc3-spa and 2.6.7-rc2-mm2 and put them up at:
http://zeke.yi.org/linux/spa/ . There is also vmstat output there but it doesn't look too helpful to me.

Thanks for the feedback, I'll add your test to those I perform myself.

Some of the control variables for this scheduler have rather arbitrary values at the moment and are likely to be non optimal. I'm in the process of making some of these variables modifiable at run time via /proc/sys to enable experimentation with different settings. Hopefully, this will enable settings that improve interactive performance to be established.

Once again, thanks for the feedback
Peter

Quoted 10 lines · Peter Williams

Quoted 8 lines · Shane Shrybman

Hi Peter,

I just started to try out your SPA scheduler patch and found that it is noticeably sluggish when resizing a mozilla window on the desktop. I have a profile of 2.6.7-rc3-spa and 2.6.7-rc2-mm2 and put them up at:
http://zeke.yi.org/linux/spa/ . There is also vmstat output there but it doesn't look too helpful to me.

[..snip..]

Quoted 2 lines

Thanks for the feedback, I'll add your test to those I perform myself.

Sure, no problem. Hope it helps.

Quoted 1 line

Some of the control variables for this scheduler have rather arbitrary values at the moment and are likely to be non optimal. I'm in the process of making some of these variables modifiable at run time via /proc/sys to enable experimentation with different settings. Hopefully, this will enable settings that improve interactive performance to be established.

Quoted 13 lines · Shane Shrybman

Quoted 10 lines · Peter Williams

Quoted 8 lines · Shane Shrybman

Hi Peter,

I just started to try out your SPA scheduler patch and found that it is noticeably sluggish when resizing a mozilla window on the desktop. I have a profile of 2.6.7-rc3-spa and 2.6.7-rc2-mm2 and put them up at:
http://zeke.yi.org/linux/spa/ . There is also vmstat output there but it doesn't look too helpful to me.

Quoted 15 lines

[..snip..]

Quoted 2 lines

Thanks for the feedback, I'll add your test to those I perform myself.

Sure, no problem. Hope it helps.

Quoted 6 lines

Some of the control variables for this scheduler have rather arbitrary values at the moment and are likely to be non optimal. I'm in the process of making some of these variables modifiable at run time via /proc/sys to enable experimentation with different settings. Hopefully, this will enable settings that improve interactive performance to be established.

A patch which adds control of the SPA scheduler via /proc/sys/kernel is available at:

<http://users.bigpond.net.au/Peter-Williams/patch-2_6_7_rc3-SPA_CNTL-v0.1>

This adds the directory /proc/sys/kernel/cpusched to the /proc directory if SYSCTL is configured into the kernel. This directory contains 6 files that allow read/write (root only for write) to 6 unsigned integer parameters that control the scheduler as follows:

time_slice -- (in msecs) controls the amount of CPU time that CPU bound tasks get before they are booted off the CPU, have their bonuses reassessed, are put back on the run queue and rescheduled. This has a default value of 100 (which is the average time slice dealt out by the original O(1) scheduler) and a maximum allowable value of 1000 (i.e. 1 second) and a minimum of 1. Making this smaller MAY help interactive response but COULD increase the context switching rate. Making this larger MAY improve overall system throughput and reduce context switching rate but MAY make interactive responsiveness worse.

base_promotion_interval -- (in msecs) controls the rate at which the promotion mechanism promotes runnable tasks. This has a default value of 55 and a maximum of UINT_MAX and a minimum of 1. The primary effect of changing this parameter is to alter the "severity" (i.e. how much more CPU access a lower "nice" will give a task) of "nice" values. The default value is such that if there are two CPU bound tasks (hard spinners) running (and time_slice has its default value of 100) and one of them has a "nice" value one less than the other then that task will get one and a half times as much CPU access as the other.

max_ai_bonus -- determines the maximum interactive bonus that a task can receive. This has a default value of 10 and a maximum value of 10.
Setting this value to zero effectively turns off the interactive bonus.

max_tpt_bonus -- determines the maximum throughput bonus that a task can receive. This has a default value of 5 and a maximum value of 9 (to keep the maximum number of priority slots <= 160 (or 5 * 32)). Setting this value to zero effectively turns off the interactive bonus.

ia_threshold -- is the sleepiness (i.e. A_s / (A_s + A_c) where A_s is the average sleep duration per scheduling cycle and A_c is the average CPU consumption per scheduling cycle) threshold (in thousandths) above which the task will be considered interactive and have its interactive bonus increased asymptotically towards the maximum. This is assessed at the end of each scheduling cycle and is only used to determine whether an increase in the bonus is warranted. This has a default value of 900 and a minimum and maximum of 0 and 1000 respectively.

cpu_hog_threshold -- is the idleness (i.e.. (A_s + A_d) / (A_s + A_d + A_c) where A_c and A_s are as before and A_d is average time spent on the run queue per cycle waiting for CPU access) threshold (in thousandths) below which a task will be considered a CPU hog and start to lose interactive bonus points (if it has any). This is assessed at the end of each scheduling cycle and at the end of each time slice (if the task uses an entire slice) and is only used to decrease the task's interactive bonus. This has a default value of 500 and a minimum and maximum of 0 and 1000 respectively.

I suspect that it is this last parameter that is the cause of the phenomena that you reported as the activity you described probably caused the X servers idleness to become smaller than 500 causing it to lose some of its interactive bonus. Could you apply this patch and try decreasing the value of this parameter to see if your problem goes away?
A clue as to how small to make it can be obtained by observing the CPU
usage rate reported by top for the X server when you are stirring it up.
If the reported usage is y% then a threshold of slightly smaller than
(1000 - y * 10) should work.

Thanks
Peter
PS I will probably change the semantics of cpu_hog_threshold in the next version to use CPU usage rate (the opposite of idleness) as the measure and reverse the test (i.e. if usage rate is greater than the threshold the task loses some of its interactive bonus). Its value (A_c / (A_c + A_s + A_d) NB this is different to the complement of sleepiness) is slightly cheaper to calculate and I also think it's a more natural way of thinking about the issue.

I am doing the same basic test as before but I should mention a few things. In mozilla I have ten tabs open, 7 of them with loaded web pages. Depending on which tab is focused the interactive responsiveness varies greatly. Performance decreases rapidly with the complexity of the web page to be rendered. I am using the Debian 4.3.0-7 20040318043201 version of XFree86 with the nv driver. This driver is incredibly hungry for CPU resources (absolute pig really).

Here is what I see:

At the default setting of 500 I see mozilla and XFree86 each at about 49% cpu (fighting for cpu), each with a priority ~30-33, with no idle time. Metacity, the window manager, also looks to be trying to get some CPU, cause it is usually third on the list at ~3% cpu.

As cpu_hog_threshold is decreased to 400. CPU X=50% pri=30, mozilla=40% pri=34-37, metacity=4.5% pri=29. Feels a tad less jerky, cause metacity gets a bit more cpu?

Decreasing cpu_hog_threshold to 300 feels worse than 500.

Decreasing the time_slice to 10-20 makes things noticeably smoother. Are time slices dynamic in the SPA scheduler?

and the vanilla (0)1
scheduler?

My humble assessment is that to provide a good interactive feel for this workload; Xfree86, mozilla and metacity all need to get CPU in a timely manner. This is a tricky balancing act because XFree86 and mozilla can easily fall into the cpu hog category, or if they are not cpu hogs they can act to starve metacity of cpu resources.

Right now I am using cpu_hog_threshold=400 and timeslice=20 and it seems to be an improvement. But this is fairly subjective of course.