On Thu, Jun 07, 2012 at 04:45:56PM +0200, Michal Hocko wrote:

On Fri 01-06-12 10:37:30, Michal Hocko wrote:
[...]

quoted

More detailed statistics (max/min - the worst/best performance).
	comparison (cong is 100%)	comparison (page reclaim 100%)			
	max	min	median		max	min	median
* ext3
** Write
5M	171.20%	95.33%	98.70%		216.96%	101.99%	103.61%
60M	97.56%	98.80%	104.51%		110.09%	100.11%	116.59%
300M	99.76%	99.49%	99.35%		99.47%	99.89%	99.57%
2G	99.52%	99.53%	99.52%		100.09%	99.07%	100.02%

** Read					
5M	35.37%	38.70%	39.09%		83.55%	89.85%	86.54%
60M	89.70%	102.90%	102.00%		97.71%	101.91%	102.06%
300M	92.38%	99.33%	99.14%		80.65%	98.39%	91.23%
2G	90.07%	99.92%	100.38%		99.85%	100.75%	99.94%

* Tmpfs					
** write
5M	121.85%	99.69%	131.57%		219.22%	99.85%	135.30%
60M	140.82%	99.70%	139.57%		98.14%	54.51%	73.65%
300M	97.99%	99.54%	99.60%		99.29%	99.57%	99.32%
2G	99.37%	99.62%	99.64%		98.72%	99.92%	99.18%

** read				
5M	85.44%	92.96%	88.92%		129.13%	101.54%	97.87%
60M	64.41%	94.35%	88.10%		97.41%	95.75%	96.31%
300M	116.89%	106.52%	120.84%		132.17%	104.39%	130.63%
2G	86.27%	99.96%	87.47%		60.69%	99.44%	98.49%

I have played with the patch below but it didn't show too much
difference in the end or we end up doing even worse. 

Here is the no_patch/patched comparison:

	comparison (page reclaim is 100%)
* ext3  avg	max	min	median
** Write
5M    	81.49%	77.53%	101.91%	76.60%
60M   	98.60%	95.58%	101.40%	99.62%
300M  	101.68%	102.05%	101.19%	101.73%
2G    	102.20%	102.25%	102.12%	102.22%
				
** Read  				
5M    	103.94%	105.14%	103.95%	103.32%
60M   	105.26%	107.91%	103.15%	104.95%
300M  	104.83%	107.86%	101.65%	104.88%
2G    	102.67%	101.26%	102.83%	103.35%

* Tmpfs
** Write
5M    	107.68%	119.66%	105.26%	102.78%
60M   	122.16%	138.51%	103.62%	121.09%
300M  	101.03%	100.67%	101.11%	101.17%
2G    	101.82%	101.66%	101.87%	101.87%
				
** Read			
5M    	102.47%	124.02%	98.05%	92.57%
60M   	103.62%	121.03%	96.97%	96.52%
300M  	98.90%	118.92%	102.64%	86.19%
2G    	83.50%	76.34%	97.36%	81.92%

I am not sure it really makes sense to play with the priority here. All
the values we would end up with would be just wild guesses or mostly
artificial workloads. So I think it makes some to go with the original
version of the PageReclaim patch without any further fiddling with the
priority.

Is this sufficient to go with the patch or do people still have concerns
which would block the patch from merging?

No, let's go for it.  It's a net improvement as it stands.

Acked-by: Johannes Weiner <hannes@cmpxchg.org>

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@kvack.org.  For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"dont@kvack.org"> email@kvack.org </a>