BFS Scheduler Benchmarks

[RealNC]

I think Michael ran the ab benchmark on the same machine as Apache. Usually you run ab from another machine and Apache on the machine that is benchmarked.

[vesa]

BFS seems to be doing something seriously right, so here is some reporting from my experiments. Pulled from mail to Con.

> First of all, thank you for your fine work on Linux scheduler. I've been
> really surprised by the low userspace latencies BFS brought. Testing so
> far has been some kernel compilation with video and audio playback, but
> most importantly all the above with timed HW I/O measurement. That means
> a digital storage oscilloscope attched to serial port RTS line and
> wiggling it up/down (40 ms period) with few lines of python. Results are
> groundbreaking The worst case latency was 10ms on saturated machine
> without any job nicing or schedtool usage. Unloaded latency was always
> below 2ms. Stock kernel would break totally. The machine is Core2 duo
> 1.86GHz with Q965 chipset and integrated graphics. Ubuntu 9.04 64b base
> system and graphics drivers from git.

The piece of python used was:

import serial, time
port = serial.Serial(1) #Serial port number
while(True):

time.sleep(0.02)
port.setRTS(0)
time.sleep(0.02)
port.setRTS(1)

You'll need to have pyserial installed and it's called python-serial in Ubuntu. The signal was taken from pins 5 (ground) and 7 (RTS) by connecting a led and 1k resistor for load. That allows convenient visual feedback too as a bonus. There are many ways of improvement for data logging and very good instrument can be built from some microcontroller like AVR/PIC/8051. Uc could be powered from the port and also upload the interval histogram data to the host system too. Current measurements were taken with old trusty Tektronix 486 with waveform envelope storage, so there is no histogram data available yet.

I've also tested RT-kernels in similar ways (RT wiki is a good start), and very interesting test would be BFS+RT. Con said that he has no time for studying the RT patch set, but is willing to help anyone taking the task. So there it is for the adventurous.

Also BFS solves some of the problems depicted here http://www.phoronix.com/forums/showthread.php?t=18108
as much of the tearing and missed syncs/frames are really sheduling issue. Our DVB here is locked to rubidium and if I set my framrate close enough to 50Hz, the field slips are so rare that there are way more glitches in low quality overseas streams. Domestic stuff is technically top quality and there one may indeed notice the error.

[tuxdriver]

It's sad, but even Win95 had very good 2D acceleration,much better than any Linux distro I've tried on this Dual-Core box. Modern Windows OSes are a lot faster. The 2D GUI performance of Windows XP is out of this world.

Linux can be just as fast: Extremely responsive system with GUI performance that rivals Windows XP.

All you need is an older machine with a Radeon R200 card (Radeon 8500 ~ Radeon 9250) and this live CD:

http://www.puppyisos.org/isos/2008-0...7-fglrx8.28.8/

username: puppy passwd: linux

Good old fglrx for R200 cards = the pinnacle of ATI's proprietary drivers for Linux

You'll be amazed.

[mo0n_sniper]

First of all I want to say that I didn't test the BFS scheduler.

Second I think that the test isn't relevant with what a desktop scheduler should do.
The current state is that my machine usually locks for 1-2 seconds when I'm doing high i/o operations.I'm preety sure it's the scheduler fault and also I have a resonably good machine(1.6 Ghz dual core,2Gb RAM).
The scheduler shouldn't allow for this lock-ups even if it would mean that the job will take .1 seconds longer.
Also I don't think that designing the scheduler for a desktop with 4000+ processors in mind is a good ideea.

[kraftman]

I am amazed at the web performance results. Could that be expanded into more real-life situations: static pages with nginx, full LAMP stack (with PHP, Perl and Python), django, turbo gears, ruby on rails, Tornado, ...

In high end web business people spend millions just to get a few percent of performance gain, if BFS or a variant of it can give a 10% speed boost in heavily loaded web serving situation, it would be a godsend and would be pushed as an option into mainline very, very rapidly.

I repeated this two times. CFS is faster with fair_sleepers disabled. There's also some objective benchmark at lkml. Read RealNC post also.

http://lkml.org/lkml/2009/9/15/75

[kraftman]

Linux can be just as fast: Extremely responsive system with GUI performance that rivals Windows XP.

All you need is an older machine with a Radeon R200 card (Radeon 8500 ~ Radeon 9250) and this live CD:

http://www.puppyisos.org/isos/2008-0...7-fglrx8.28.8/

username: puppy passwd: linux

Good old fglrx for R200 cards = the pinnacle of ATI's proprietary drivers for Linux

You'll be amazed.

Thanks Tuxdriver So, weak 2D is only drivers problem and we can have superb 2D acceleration without Gallium3D? (In theory )

P.S. I have r500, but it won't be so smooth with this card?

[kraftman]

First of all I want to say that I didn't test the BFS scheduler.

I recommend you to test it, because BFS had problems here I mentioned before (huge game slowdowns when listening to the music same time and dead keyboard and mouse input which happen from time to time). The problem you described is probably related to something else - bug, but not in CFS, firmware etc. I had similar problem, but system hung for more then ten seconds sometimes. It seems to be fixed with 2.6.30.6 but I didn't test it much. However, if I trigger this bug I'll check if the same will happen with BFS.

Btw. Con knows about such problems I described

[kraftman]

@Vesa

Have you tried with fair_sleepers disabled? :> It makes a huge difference sometimes. It would be great if you also post this at lkml.

[Solitary]

CFS is faster with fair_sleepers disabled. At least here.

According to http://lkml.org/lkml/2009/5/30/84. Doesn't "fair_sleepers disabled" increase throughput ie batch processing but also increase latency?

Seems that would make system responsiveness less than before. A tweak for servers, not desktops?

[kraftman]

According to http://lkml.org/lkml/2009/5/30/84. Doesn't "fair_sleepers disabled" increase throughput ie batch processing but also increase latency?

Seems that would make system responsiveness less than before. A tweak for servers, not desktops?

It probably can have different effects on different configurations. It solved some people problems with latency:

http://lkml.org/lkml/2009/9/10/229

I can't say a word about this, because my system is responsive in both cases.