2.6.14-ck4 (staircase v13)

[Con Kolivas]

These are patches designed to improve system responsiveness and interactivity. 
It is configurable to any workload but the default ck patch is aimed at the 
desktop and cks is available with more emphasis on serverspace.

Apply to 2.6.14
http://ck.kolivas.org/patches/2.6/2.6.14/2.6.14-ck4/patch-2.6.14-ck4.bz2

or server version
http://ck.kolivas.org/patches/cks/patch-2.6.14-cks4.bz2

web:
http://kernel.kolivas.org
all patches:
http://ck.kolivas.org/patches/
Split patches available.


Changes:

Added:
 +patch-2.6.14.1
Latest stable patch

 +net-fix_zero-size_datagram_reception.patch
Bugfix queued for 2.6.14.2

 +sched-staircase12.2_13.patch
Major update in the staircase code. This now makes the interactive mode work 
out a simple mathematical calculation to determine the cpu percentage of 
entitled cpu a task uses and use that to determine what best dynamic priority 
(and thus how low the latency is). This means that a low cpu using task can 
have very low latency even in the presence of much less niced tasks. For 
example a nice 0 task in the presence of a nice -20 task can still have low 
latency if it uses very little cpu. This means that 'nice' is a determinant 
of cpu entitlement, but no longer the prime determinant of latency. It is 
thus possible to even run low cpu using audio apps at nice 19. Running audio 
at nice 19 would be unusual, though. However this does help for the opposite 
as well - there are kernel threads that run nice 0, -5 and even -20, when 
normal user tasks run at nice 0. Some of these kernel threads and workqueues 
can use extraordinary amounts of cpu as well. These are real world examples 
where it would help on normal desktops now. Interbench benchmarks demonstrate 
this below.

 -2614ck3-version.diff
 +2614ck4-version.diff
Version update


---


If you are new to interbench results, the order of importance is from right to 
left, where deadlines met and desired cpu should be as close to 100% as 
possible, amd max latency and average latency as low as possible. 


Asterisks have been placed to the right to indicate where one kernel 
outperforms the other. Use a fixed font to be able to read this table!

Interbench 0.29 results on 3Ghz P4 uniprocessor:

Here is mainline 2.6.14, with the benchmarked interactive process running at 
nice 19:

--- Benchmarking simulated cpu of Audio nice 19 in the presence of simulated
Load	Latency +/- SD (ms)  Max Latency   % Desired CPU  % Deadlines Met
None	  0.004 +/- 0.00442    0.008		 100	        100
Video	  0.004 +/- 0.00491    0.009		 100	        100
X	   36.4 +/- 62.7         233		  96	       63.8
Burn	    673 +/- 699          701		0.568	      0.568
Write	   1.39 +/- 9.22         118		99.5	       99.5
Read	   0.01 +/- 0.0133     0.105		 100	        100
Compile	    774 +/- 803         1003		0.559	          0
Memload	   4.78 +/- 27.9         225		95.5	       95.5

--- Benchmarking simulated cpu of Video nice 19 in the presence of simulated 
Load	Latency +/- SD (ms)  Max Latency   % Desired CPU  % Deadlines Met
None	  0.004 +/- 0.00493    0.026		 100	        100	*
X	   21.9 +/- 50           266		54.2	       48.8
Burn	    741 +/- 770          771		0.174	      0.174
Write	   2.58 +/- 11.2         140		96.1	       90.1
Read	  0.009 +/- 0.0176     0.182		 100	        100
Compile	    758 +/- 783          877		0.173	          0
Memload	   2.38 +/- 20.3         269		93.9	       92.9

--- Benchmarking simulated cpu of X nice 19 in the presence of simulated ---
Load	Latency +/- SD (ms)  Max Latency   % Desired CPU  % Deadlines Met
None	  0.009 +/- 0.1            1		 100	         99	*
Video	     13 +/- 23.7          60		69.8	         53
Burn	    607 +/- 634          750		1.15	       1.15
Write	   6.48 +/- 17.4          90		80.5	       66.7
Read	  0.227 +/- 1.05           6		92.7	       89.6
Compile	    680 +/- 720          957		1.14	       1.14
Memload	   7.84 +/- 31.7         272		76.6	       66.4

Note the massive dropped deadlines even with something using just 5% cpu when 
run at nice 19


Here is 2.6.14-ck4 running staircase v13:

--- Benchmarking simulated cpu of Audio nice 19 in the presence of simulated 
Load	Latency +/- SD (ms)  Max Latency   % Desired CPU  % Deadlines Met
None	  0.004 +/- 0.00419    0.009		 100	        100
Video	  0.004 +/- 0.00457    0.008		 100	        100
X	  0.054 +/- 0.388          4		 100	        100	*
Burn	   3.27 +/- 18.8         136		  98	       97.4	*
Write	    1.3 +/- 11.7         120		  99	         99	*
Read	  0.011 +/- 0.0119     0.056		 100	        100
Compile	   2.81 +/- 16.6         126		  98	         98	*
Memload	   3.39 +/- 22           206		97.5	       97.5	*

--- Benchmarking simulated cpu of Video nice 19 in the presence of simulated 
Load	Latency +/- SD (ms)  Max Latency   % Desired CPU  % Deadlines Met
None	  0.031 +/- 0.682       16.7		 100	       99.8
X	   9.87 +/- 25.5         133		74.8	       67.4	*
Burn	   6.33 +/- 31.7         483		85.8	       79.6	*
Write	   1.69 +/- 11.7         172		97.1	       94.6	*
Read	  0.011 +/- 0.0142     0.107		 100	        100
Compile	   7.17 +/- 46.4         617		78.3	       75.7	*
Memload	   2.46 +/- 20.8         245		92.9	       92.9

--- Benchmarking simulated cpu of X nice 19 in the presence of simulated ---
Load	Latency +/- SD (ms)  Max Latency   % Desired CPU  % Deadlines Met
None	  0.237 +/- 2.3           23		 100	         98
Video	     13 +/- 23.6          60		69.8	         53
Burn	    327 +/- 659         1398		20.5	       16.4	*
Write	    2.9 +/- 13.4         108		73.3	       69.9	*
Read	  0.217 +/- 1.01           6		92.7	       89.6
Compile	    365 +/- 691         1519		19.3	       14.3	*
Memload	   8.42 +/- 19.8          92		80.8	       73.2	*


You can see here that the audio benchmark meets more than 97% of the deadlines 
even in extreme loads when run at nice 19, and interbench ignores the effect 
of buffering. With some buffering it is unlikely to skip sound even at nice 
19. Now if you move on to video and then X you can see that the improvement 
over mainline gets progressively less as each load uses more cpu. This is 
because the latency is determined by cpu% of entitlement. Even so, the 
deadlines met are significantly higher (especially for video which doesnt 
drop below 67% compared to mainline which bottoms out at 1%). 


---


Moving on to the more "real world" example of nice 0 interactive task in the 
presence of background loads running at nice -5:

Mainline 2.6.14:
--- Benchmarking simulated cpu of Audio in the presence of simulated nice -5 
Load	Latency +/- SD (ms)  Max Latency   % Desired CPU  % Deadlines Met
None	  0.004 +/- 0.00526    0.029		 100	        100
Video	  0.004 +/- 0.00484    0.007		 100	        100
X	     36 +/- 59.8         113		75.5	       63.3
Burn	   6.78 +/- 90.5        1200		88.1	         88
Write	   2.21 +/- 13.5         141		  99	         99
Read	   0.01 +/- 0.0146     0.111		 100	        100
Compile	   22.1 +/- 183         1600		69.7	       69.3
Memload	  0.191 +/- 0.957       7.07		 100	        100

--- Benchmarking simulated cpu of Video in the presence of simulated nice -5 
Load	Latency +/- SD (ms)  Max Latency   % Desired CPU  % Deadlines Met
None	  0.004 +/- 0.00415     0.01		 100	        100	*
X	   21.4 +/- 51.1         417		53.7	       48.2
Burn	   23.6 +/- 186         1467		41.4	         41
Write	   3.11 +/- 16.9         315		93.7	       88.8
Read	  0.009 +/- 0.0145      0.12		 100	        100
Compile	   21.5 +/- 171         1400		43.7	       43.1
Memload	  0.292 +/- 2.08        28.3		 100	         99

--- Benchmarking simulated cpu of X in the presence of simulated nice -5 ---
Load	Latency +/- SD (ms)  Max Latency   % Desired CPU  % Deadlines Met
None	  0.009 +/- 0.0995         1		 100	         99
Video	     13 +/- 23.7          60		69.8	         53	*
Burn	   77.5 +/- 306         1404		45.6	       43.8
Write	   9.71 +/- 28.4         224		76.2	       64.9
Read	  0.245 +/- 1.13           6		91.1	       88.9
Compile	   97.6 +/- 365         1564		41.1	       39.2
Memload	   4.31 +/- 12.7          72		57.6	       51.2

You can see that without any buffering, it is possible even at nice 0 for 
audio to miss almost 40% of the deadlines under a nice -5 X load (or 
equivalent load from kernel threads etc).


And here is 2.6.14-ck4:

--- Benchmarking simulated cpu of Audio in the presence of simulated nice -5 
Load	Latency +/- SD (ms)  Max Latency   % Desired CPU  % Deadlines Met
None	  0.003 +/- 0.00351    0.009		 100	        100
Video	  0.004 +/- 0.00451    0.009		 100	        100
X	  0.287 +/- 1.43          10		 100	        100	*
Burn	  0.004 +/- 0.00457    0.008		 100	        100	*
Write	  0.009 +/- 0.0102     0.018		 100	        100	*
Read	  0.011 +/- 0.0128     0.069		 100	        100
Compile	  0.009 +/- 0.0107     0.039		 100	        100	*
Memload	  0.037 +/- 0.167          2		 100	        100

--- Benchmarking simulated cpu of Video in the presence of simulated nice -5 
Load	Latency +/- SD (ms)  Max Latency   % Desired CPU  % Deadlines Met
None	  0.031 +/- 0.682       16.7		 100	       99.8
X	   3.43 +/- 8.46        27.7		 100	         83	*
Burn	  0.032 +/- 0.681       16.7		 100	       99.8	*
Write	  0.006 +/- 0.00846     0.07		 100	        100	*
Read	  0.011 +/- 0.0115      0.04		 100	        100
Compile	   0.04 +/- 0.683       16.7		 100	       99.8	*
Memload	  0.062 +/- 0.694       16.7		 100	       99.8	*

--- Benchmarking simulated cpu of X in the presence of simulated nice -5 ---
Load	Latency +/- SD (ms)  Max Latency   % Desired CPU  % Deadlines Met
None	  0.009 +/- 0.0995         1		 100	         99
Video	   13.5 +/- 24.1          60		67.7	       50.8
Burn	  0.117 +/- 0.783          6		98.1	       96.1	*
Write	  0.147 +/- 0.751          4		94.4	       92.4	*
Read	  0.009 +/- 0.0995         1		 100	         99	*
Compile	  0.147 +/- 0.751          4		94.4	       92.4	*
Memload	  0.176 +/- 1.69          17		 100	         99	*

Whereas here audio does not drop a single deadline even without any buffering. 
Video and even X also show significant improvement.


Cheers,
Con

Attachment: pgp00000.pgp
Description: PGP signature