Re: strange raid5

[Ming Zhang <mingz@ele.uri.edu>]

On Sun, 2005-07-24 at 07:37 +1000, Neil Brown wrote:
> On Saturday July 23, mingz@ele.uri.edu wrote:
> > 1048576 = 1024 * 1024 = 32 * 32768. :)
>       ^^^
> > 
> > so it should be 32 stripe writes.
> > 
> > ming
> > 
> > On Fri, 2005-07-22 at 23:14 -0700, Tyler wrote:
> > > By my calculations, 1048756 is *not* a multiple of 32768 (32 
>                             ^^^
> > > Kilobytes).  Did I miscalculate?
> > > 
> 
> A typo somewhere :-)
yes, my dumb stupidness. a typo here.

@Tyler, sorry about this. :P


i checked again and what i did is 
./write /dev/md0 1048576 1024 s

and i still see plenty read.

i build raid with this no resync. but should be ok, right?

mkraid -c raidtab -R --dangerous-no-resync /dev/md0

my raidtab file is like this

raiddev                 /dev/md0
raid-level              5
nr-raid-disks           3
chunk-size              32
parity-algorithm        left-symmetric

device                  /dev/sda
raid-disk               0

device                  /dev/sdb
raid-disk               1

device                  /dev/sdc
raid-disk               2


> 
> 
> > > Regards,
> > > Tyler.
> > > 
> > > Ming Zhang wrote:
> > > 
> > > >i created a 32KB chunk size 3 disk raid5. then write this disk with a
> > > >small code i wrote. i found that even i write it with 1048756 in unit,
> > > >which is multiple of stripe size, it still has a lot of read when seen
> > > >from iostat. 
> 
> > > >sda             605.05       387.88     35143.43        384      34792
> > > >sdb             611.11       323.23     35143.43        320      34792
> > > >sdc             602.02       387.88     35143.43        384      34792
> 
> I wouldn't call this "a lot of read".  The read requests are only 1%
> of the write requests.  So I would call it "some read".
> 
> There is quite a lot of complexity between the 'write' system all and
> the data actually getting to the device.  Presumably the Linux VM
> system is flushing dirty data to the device at times other than then
> end of the write request.
> I think the block layer also automatically flushes devices every
> 200msecs.
> This may be triggering flush requests which aren't stripe-aligned.
> 
> You are certainly getting the vast majority of stripes written as
> whole stripes.
> 
> NeilBrown