Re: [PATCH] sata_sil: Greatly improve DMA support

[Tejun Heo <htejun@gmail.com>]

Indan Zupancic wrote:
> On Sun, May 20, 2007 00:03, Jeff Garzik wrote:
>> Indan Zupancic wrote:
>>> This patch seems to work with my SiI 3512, though I don't notice any
>>> difference, neither a speedup, nor a slowdown. Hdparm gives the same
>>> speeds (-tT), and cp -a'ing kernel sources is abysmal slow in both cases,
>>> (need to look into that one) so I didn't really test it that well.
>>
>> It won't result in much of a speedup, except in situations where IOMMU
>> or other situation that causes you to run into the 64k boundary being an
>> issue -- generally only on huge transfers.
>>
>> A good measure is to dd(1) to/from the block device, rather than using a
>> filesystem.  As has been shown on LKML, the filesystem can really slow
>> things down in some cases.
> 
> I didn't really expect a speedup, it's more that I've no regression to report.
> 
> I could benchmark the patch more thoroughly, but right now I'm more worried
> about the crawling cp I just discovered. Talking about filesystems slowing down
> things...
> 
> Test:
> 
> $ cp -a linux-2.6/ /tmp/
> 
> done on the same ext3 partition. linux-2.6 contains source and git repo only,
> I'm compiling stuff with O=../obj.
> 
> $ vmstat 10
> procs -----------memory---------- ---swap-- -----io---- -system-- ----cpu----
>  r  b   swpd   free   buff  cache   si   so    bi    bo   in   cs us sy id wa
>  0  1      0   4168   3316 195700    0    0   739   494  530  393 15  3 66 16
>  0  3      4   4120   2040 198196    0    0 14677 14111 1247  435  0 17  0 83
>  0  1      4   3588   1444 199696    0    0  8892  9472 1362  438  0 12  0 88
>  1  0      4   3772   4228 196012    0    0   764   454 1161  345  0  4  0 96
>  0  1      4   3548   6156 193088    0    0   793   851 1158  340  0  4  0 96
>  0  1      4   3852   7608 189096    0    0   798   523 1160  474  1  4  0 95
>  1  1      4   3612   8684 186048    0    0  1244   864 1178  430  2  5  0 93
>  0  1      4  90660   9308  96396    0    0   853   906 1244  578  7  6  0 87
>  0  1      4  72280   9816 112368    0    0   830   854 1278  429 12  5  0 83
>  1  0      4  52488  10296 130560    0    0   935   861 1178  418  1  6  0 94
>  0  1      4  30500  10788 149776    0    0   977   858 1178  371  0  6  0 94
>  0  1      4   9792  11244 167856    0    0   918  1394 1182  350  1  5  0 94
>  0  1      4   4016  11216 172504    0    0  1017   858 1181  382  1  6  0 94
>  0  1      4   3660  11484 171484    0    0   966   861 1182  410  1  6  0 94
> 
> It never finished, as I had no patience to copy about 900 Mb with this rate.
> 
> As it's a git tree, I suppose it's heavily fragmented, but this is still rather
> pathetic. Should I blame cp, or is something else wrong? Any ideas how
> to figure this one out would be appreciated. Sorry for the off-topicness.

Do things improve if you change the io scheduler to deadline?

  # echo deadline > /sys/block/sda/queue/scheduler

Also worth looking at is the following bug entry.

  http://bugzilla.kernel.org/show_bug.cgi?id=7372

There seems to be weird interaction among the scheduler / VM / IO.  The
exact cause is still not verified.  :-(

-- 
tejun