List:General Discussion« Previous MessageNext Message »
From:Alec.Cawley Date:August 26 2003 8:51am
Subject:Re: RAID or not?
View as plain text  

> Alec.Cawley@stripped wrote:


> > I do not think it is true that mirroring gives no performance benefit
(on a
> > well implemented controller). For reads, the raid controller can read
> > either copy of the data, so that effectively two reads can be in
progress
> > at the same time, doubling read performance.

> Yes, doubling read performance (best case) when compared with a non-RAID
disk
> system. But with RAID5, you have more disks and can therefore read more
than
> just two disks simultaneously.  The more spindles you have, the better
the
> performance.

I was assuming that your number of spindles was set by your need to store
daya. If this is so, the number of primary data drives, and hence spindles,
is fixed. It is noly whtehr you have one parity drive, or N mirror drives.
The latter is, of course, more expensive for a given storage capacity - but
gives you higer performance for your money.

> > On the write side, for small
> > writes a raid 5 has to read the overwritten data (in order to remove it
> > from the parity) then do a read/modify/write on the parity.

> I don't quite understand the need to read data before any write.  Why
wouldn't
> it just calculate the parity of whatever is being written and just write
it to
> disk?  Wouldn't there be slack space, as with any disk system?  Write a 1
byte
> file and it uses an N byte block on one disk, plus an N byte parity block
on
> another.

Call the disks A, B, C, D, P. For some sector we are about to overwrite A
with A'. The current contents of P are A^B^C^D (^ used for XOR). The final
contents of P must be A'^B^C^D. We can read B, D and D to produce this
valure ab initio. However, it is more efficient to read A and P, XOR them
to produce B^C^D then XOR in the new data A'. In this case, both A and P
must do read followed by write. Since this involves a full rotation between
the read and the write, this is a relatively slow operation. This does not
occur if we are overwrting all of A, B, C, and D because *in this case* we
can do as you expect and calculate P on the fly.

> > Performance
> > again should be doubled (two writes on the mirrored system,
read/overwrite
> > and read/modify/write on the Raid5 system, with the two halves of each
> > operation requiring a full rotation between them). For large writes,
the
> > Raid5 system catches up, because the parity can be entirely calculated
from
> > the data sent, so it does not need to do the reads.


> > I would therefore expect a mirrored system to approach, but not reach,
> > twice the performance of a Raid5 system

> A mirrored system when writing should, at best, be no faster than a
> non-mirrored system.  Unlike reading, it has to write a complete copy of
the
> data to both disks, so there are no performance gains to be had.  RAID5
should
> be a bit slower than RAID1, due to the need for calculating parity.

See explanation above. For large writes, you are correct. For smaller ones,
and for the scruffy bits at the beginning and end of large writes, RAID 5
performance will approach 50% of Raid 1.

      Alec




Thread
RAID or not?Jackson Miller21 Aug
  • Re: RAID or not?Jon Drukman21 Aug
    • Re: RAID or not?Jackson Miller21 Aug
      • RE: RAID or not?Steven) (Lefevre21 Aug
        • Re: RAID or not?David Griffiths21 Aug
        • Re: RAID or not?Per Andreas Buer21 Aug
        • Re: RAID or not?Jon Drukman22 Aug
          • Re: RAID or not?Dan Nelson22 Aug
          • RE: RAID or not?Steven) (Lefevre22 Aug
            • RE: RAID or not?Michael Loftis2 Sep
  • Re: RAID or not?Colbey21 Aug
  • Re: RAID or not?Per Andreas Buer21 Aug
RE: RAID or not?Alec.Cawley22 Aug
  • Re: RAID or not?Jim McAtee23 Aug
    • Re: RAID or not?Michael Loftis2 Sep
RE: RAID or not?Alec.Cawley22 Aug
Re: RAID or not?Alec.Cawley26 Aug