Re: Raid 10 Problems?

Re: Raid 10 Problems?

[Jan Engelhardt]

On Mar 7 2007 10:20, dean gaudet wrote:
>>> http://gentoo-wiki.com/HOWTO_Install_on_Software_RAID#Write-intent_bitmap
>> 
>> That information has been extremely useful. Thanks a
>> lot. I fund a command to do the bitmap internal after
>> the array was made so I added that. Seems like some of
>> these features should be default. Maybe it's time for
>> the raid folks to update what is default?
>
>the bitmap has performance implications... for example:
>http://www.mail-archive.com/linux-raid@vger.kernel.org/msg07229.html

I wonder if bitmapping a raid1 volume is faster than bmp.ing raid5.

The other thing is, the bitmap is supposed to be written out at intervals,
not at every write, so the extra head movement for bitmap updates should
be really low, and not making the tar -xjf process slower by half a
minute.
Is there a way to tweak the write-bitmap-to-disk interval? Perhaps 
something in /sys or ye olde /proc. Maybe linux-raid@ knows 8)


>note that unless you tweak your init scripts you'll need to put external 
>bitmaps on your root partition, see this thread:

Huh? That statement does not make sense. But I think you meant: when using
external bitmaps, adjust the init scripts. Because internal bitmaps are good
for one thing: you don't need to change anything.


Jan
-- 

Re: Raid 10 Problems?

[Michael Tokarev]

Jan Engelhardt wrote:
[]
> The other thing is, the bitmap is supposed to be written out at intervals,
> not at every write, so the extra head movement for bitmap updates should
> be really low, and not making the tar -xjf process slower by half a minute.
> Is there a way to tweak the write-bitmap-to-disk interval? Perhaps 
> something in /sys or ye olde /proc. Maybe linux-raid@ knows 8)

Hmm.  Bitmap is supposed to be written before actual data write, to mark
the to-be-written areas of the array as "being written", so that those
areas can be detected and recovered in case of power failure during
actual write.

So in case of writing to a clean array, head movement always takes place -
first got to bitmap area, and second to the actual data area.

That "written at intervals" is about clearing the bitmaps after some idle
time.

In other words, dirtying bitmap bits occurs right before actual write,
and clearing bits occurs at intervals.

Sure, if you write to (or near) the same place again and again, without
giving a chance to md subsystem to actually clean the bitmap, there will
be no additional head movement.  And that means, for example, tar -xjf
sometimes, since filesystem will place the files being extracted close to
each other, thus hitting the same bit in the bitmap, hence md will skip
repeated bitmap updates in this case.

/mjt

Re: Raid 10 Problems?

[Marc Perkel]

--- Michael Tokarev <mjt@tls.msk.ru> wrote:

> Jan Engelhardt wrote:
> []
> > The other thing is, the bitmap is supposed to be
> written out at intervals,
> > not at every write, so the extra head movement for
> bitmap updates should
> > be really low, and not making the tar -xjf process
> slower by half a minute.
> > Is there a way to tweak the write-bitmap-to-disk
> interval? Perhaps 
> > something in /sys or ye olde /proc. Maybe
> linux-raid@ knows 8)
> 
> Hmm.  Bitmap is supposed to be written before actual
> data write, to mark
> the to-be-written areas of the array as "being
> written", so that those
> areas can be detected and recovered in case of power
> failure during
> actual write.
> 
> So in case of writing to a clean array, head
> movement always takes place -
> first got to bitmap area, and second to the actual
> data area.
> 
> That "written at intervals" is about clearing the
> bitmaps after some idle
> time.
> 
> In other words, dirtying bitmap bits occurs right
> before actual write,
> and clearing bits occurs at intervals.
> 
> Sure, if you write to (or near) the same place again
> and again, without
> giving a chance to md subsystem to actually clean
> the bitmap, there will
> be no additional head movement.  And that means, for
> example, tar -xjf
> sometimes, since filesystem will place the files
> being extracted close to
> each other, thus hitting the same bit in the bitmap,
> hence md will skip
> repeated bitmap updates in this case.
> 
> /mjt
> 

I assume that if a block is already dirty then that is
cached somewhere in memory so you aren't writing to
the bitmap unless you're changing it for clean to
dirty? If that's the case then I would think that
writing to the map wouldn't be that expensive?



 
____________________________________________________________________________________
Now that's room service!  Choose from over 150,000 hotels
in 45,000 destinations on Yahoo! Travel to find your fit.
http://farechase.yahoo.com/promo-generic-14795097