Re: [PATCH 0/8][RFC] IO latency/throughput fixes

[Linus Torvalds]

On Mon, 6 Apr 2009, Linus Torvalds wrote:
> 
> Yes. I want to go back and see how much this all helps for the 
> data=ordered case (if at all), but I think we should at least consider 
> trying 'data=writeback' as a default.

Hmm.

So I'm back to "data=ordered", and quite frankly, so far it's a horrible 
experience. 

The difference is rather startlingly huge, which it was _not_ before. 

Sure, data=writeback was better before too, but in the end, the difference 
between the occasional 5+ second pause and the occasional 10+ second pause 
wasn't really all that interesting. They were both unusuable, and both 
made me kill the background writer almost immediately.

This time I really forced myself to keep it around, just to get the whole 
horridness of the experience.

Here's wahat I just got from data=writeback and all the latest patches 
(which includes one extra experimental one from Jens to further improve on 
things):

        fsync(6)                                = 0 <0.839397>
        fsync(6)                                = 0 <0.879706>
        fsync(6)                                = 0 <0.240856>
        fsync(6)                                = 0 <0.006497>
        fsync(6)                                = 0 <0.207197>
        fsync(4)                                = 0 <0.318396>
        fsync(4)                                = 0 <0.044499>
        fsync(4)                                = 0 <0.088381>
        fsync(6)                                = 0 <0.001057>
        fsync(9)                                = 0 <0.245389>
        fsync(7)                                = 0 <0.042728>
        fsync(7)                                = 0 <1.148262>
        fsync(6)                                = 0 <0.200707>
        fsync(6)                                = 0 <0.045259>
        fsync(6)                                = 0 <0.111846>

that's all quite usable. Stuttering, yes, but never anything feeling rally 
bad, or like something died.

This is with data=ordered on the same kernel:

	fsync(5)                                = 0 <0.000949>
	fsync(8)                                = 0 <19.674681>
	fsync(8)                                = 0 <1.169357>
	fsync(8)                                = 0 <17.994631>
	fsync(8)                                = 0 <5.711143>
	fsync(8)                                = 0 <4.759515>
	fsync(4)                                = 0 <17.283171>
	fsync(4)                                = 0 <4.371930>
	fsync(4)                                = 0 <0.008327>
	fsync(8)                                = 0 <0.000636>
	fsync(11)                               = 0 <9.802027>
	fsync(9)                                = 0 <14.109262>
	fsync(9)                                = 0 <4.081919>
	fsync(8)                                = 0 <0.000490>
	fsync(8)                                = 0 <0.001588>
	fsync(11)                               = 0 <18.822826>
	fsync(9)                                = 0 <3.917717>
	fsync(9)                                = 0 <0.001035>
	fsync(8)                                = 0 <0.000138>
	fsync(9)                                = 0 <17.466156>

where basically each 'sync' on the big file triggers that 15-20s total 
failure. I do wonder if it's actually gotten _worse_ for the data=ordered 
case, but at the same time it does make a pretty compelling argument for 
switching the defaults around.

			Linus
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/