about the speed

about the speed

[Farkas Levente]

hi,
I'm just read
http://www.hpl.hp.com/techreports/2002/HPL-2002-352.pdf
may be it's worth to read. an intersting thing about performance (XP, 
their user-space raid and than comes kernel raid).

-- 
   Levente                               "Si vis pacem para bellum!"

Re: about the speed

[Illtud Daniel]

Farkas Levente wrote:

> I'm just read
> http://www.hpl.hp.com/techreports/2002/HPL-2002-352.pdf
> may be it's worth to read. an intersting thing about performance (XP,
> their user-space raid and than comes kernel raid).

Any comments from md developers? This is from the abstract:

 The paper contains preliminary results from an investigation comparing
 Linux and WindowsXP disk I/O using kernel-based software RAID on
 identical hardware. WindowsXP performance is shown to be substantially
 superior to Linux performance, for reasons not currently understood.

And the meat of the report:

 5.5. Direct comparison

 Of crucial interest is the direct comparison between WindowsXP,
 the kernel-based Linux software RAID implementation, and the
 userlevel software RAID implementation.

 From Figure 8 it can be seen that csraid generally outperforms
 the kernel-based RAID driver, and its performance is far more
 stable, particularly for large requests. Also, the WindowsXP
 RAID driver is astonishingly good, even compared to JBOD, with
 JBOD’s performing better for request sizes smaller than 4MB
 and WindowsXP performing better at the larger sizes.

 csraid’s performance is generally less than WindowsXP, but
 overall the loss is not catastrophic, and is likely
 sufficient for the most purposes.


Figure 8 shows that XP's RAID can be 60+% faster than md at
writing, and up to 100% faster at reading (at request size
7MB, which is rather a blip). Any inital thoughts on this report?
Has the performance changed substantially since? The distro was
Mandrake 8.2, the kernel 2.4.18-6mdk. Any comments on the
methodology or conclusions?

-- 
Illtud Daniel                                 illtud.daniel@llgc.org.uk
Uwch Ddadansoddwr Systemau                       Senior Systems Analyst
Llyfrgell Genedlaethol Cymru                  National Library of Wales
Yn siarad drosof fy hun, nid LlGC   -  Speaking personally, not for NLW
-
To unsubscribe from this list: send the line "unsubscribe linux-raid" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

Re: about the speed / XP RAID vs md

[Neil Brown]

On Monday March 31, lfarkas@bnap.hu wrote:
> hi,
> I'm just read
> http://www.hpl.hp.com/techreports/2002/HPL-2002-352.pdf
> may be it's worth to read. an intersting thing about performance (XP, 
> their user-space raid and than comes kernel raid).

It seems from the article, though they don't explicitly say, that they
are doing their timing by openning  /dev/mdX, and reading/writing on
that.
I have found myself that that isn't very fast on Linux.
Creating a filesystem on the device, and accessing a file on that
filesystem is much faster despite the filesysterm overhead.

And for myself, I am much more interested in performance of a
filesystem than performance of direct IO on /dev/mdX


Also, the raid0 layer in Linux is very thin.  There is now way it can
be adding noticable overhead.  If there are any performance problems
(and there probably are), they will be in either the block device
layer or the memory management layer.
And indeed, lots of work has been done in both of these layers in 2.5
so I have no doubt that 2.6 will get very close to raw device speeds
for raid0.

I think the results are very preliminary and that the authors need to
do a lot mo reexploring to understand exactly what is going on, as
indeed that suggest they will do.

NeilBrown