Re: Soft raid 10

"Steven Kurylo" writes:
> > Further evidence to dispel the myth that RAID is a performance solution.

RAID *may* be a performance solution, if properly implemented ;)

> I'll be the last one to say I really know what all these numbers mean,
> but the bonnie man pages says higher numbers are better.
>
> So in almost every test it was raid 10, the raid10 driver, then a single disk
> .
>
> Now RAID 5 will always be quite slow

hehe... I love generalizations like this... Please define "quite
slow" ;)

The following is an (albeit old) output from bonnie on a 9-disk RAID 5
set... using software RAID on a certain other non-Linux free OS...

              -------Sequential Output-------- ---Sequential Input-- --Random--
              -Per Char- --Block--- -Rewrite-- -Per Char- --Block--- --Seeks---
Machine MB K/sec %CPU K/sec %CPU K/sec %CPU K/sec %CPU K/sec %CPU /sec %CPU
         1500 81770 74.2 99138 52.4 10138 8.3 81102 72.7 101633 36.6 344.4 4.5

The disks in question were good for about 60MB/sec reads/writes, but
were bottlenecked by the fact that IO's were limited to 64K/IO in the
OS, and each disk was getting fed only 8KB at a time... (i.e. the
total size of each stripe was a mere 64K :( ) The disks (and the machine)
had no problem saturating the U320 SCSI bus on the system...

By reading/writing 32K/disk I could get:

              -------Sequential Output-------- ---Sequential Input-- --Random--
              -Per Char- --Block--- -Rewrite-- -Per Char- --Block--- --Seeks---
Machine MB K/sec %CPU K/sec %CPU K/sec %CPU K/sec %CPU K/sec %CPU /sec %CPU
         1500 11558 7.2 11521 2.0 9313 2.9 109517 75.6 159993 30.6 961.6 4.6

but, as you can see, you pay the huge "small write" penalty for
writes, while reads scream along quite nicely.

For just 5 of the above disks, with RAID 5, and a tuned filesystem, I
saw:

              -------Sequential Output-------- ---Sequential Input-- --Random--
              -Per Char- --Block--- -Rewrite-- -Per Char- --Block--- --Seeks---
Machine MB K/sec %CPU K/sec %CPU K/sec %CPU K/sec %CPU K/sec %CPU /sec %CPU
         1500 52633 32.1 52679 15.3 7433 2.3 107978 79.1 127167 28.0 433.1 2.8

which isn't *too* bad... But just cause someone says it's RAID 5
doesn't automatically mean it's slow :) (well... there are certain
software RAIDs and hardware RAIDs that are, but it's not true in
general...)

> and I generally would never use
> it unless disk space was more important than performance.
> Infrequently used archive storage maybe?

RAID 5 reads can be light-years faster than RAID 1 reads. So if your
data usage mix consists of mostly reads, you shouldn't discount
RAID 5... This becomes even more true when your IO mix is heavily
multi-threaded and where "more spindles" is (almost) always better...

Later...

Greg Oster
Received on Fri Jan 26 16:49:57 2007