EMC DMX Parity Raid

EMC was in yesterday giving us a the highlights of their DMX Symm
line. They were talking about it's support for "Parity Raid". They
indicated that this has been very well received and that many
companies are using it, both in it's 3+1 and 7+1 forms. Apparently,
making "Parity Raid" work well was a major design goal of the DMX.

To Quote EMC Info:
"DMX Parity RAID are now equal or better than Symmetrix 8000
mirrored configurations at the same "capacities"."


At first we thought "Parity Raid" was more like other disk systems
Raid5 implimentations, but after a few minutes it became clear that
Parity Raid is nothing more than the old Raid-S.

We have used Raid-S in the past for systems with low I/O loads, but
not for any kind of general use where I/O performance was any kind of
a consideration.

Questions:

Is the new Parity Raid really better than the old Raid-S?
Are you using Parity Raid, and if so, what for?

Thanks

Richard Rhodes

richard rhodes wrote:
To Quote EMC Info:
"DMX Parity RAID are now equal or better than Symmetrix 8000
mirrored configurations at the same "capacities"."

At first we thought "Parity Raid" was more like other disk systems
Raid5 implimentations, but after a few minutes it became clear that
Parity Raid is nothing more than the old Raid-S.

Probably just means they have put a bigger cache in front of it.

RAID-S is fine as long as you don't need to go to the physical disk -
especially for writes. As long as you're playing around in the cache
performance is fine, but the moment you hit the "wall" expect write
performance to seriously take a dive.

Scott

Scott Howard writes:

richard rhodes wrote:
To Quote EMC Info:
"DMX Parity RAID are now equal or better than Symmetrix 8000
mirrored configurations at the same "capacities"."

At first we thought "Parity Raid" was more like other disk systems
Raid5 implimentations, but after a few minutes it became clear that
Parity Raid is nothing more than the old Raid-S.

Probably just means they have put a bigger cache in front of it.

RAID-S is fine as long as you don't need to go to the physical disk -
especially for writes. As long as you're playing around in the cache
performance is fine, but the moment you hit the "wall" expect write
performance to seriously take a dive.


There really isn't any reason for RAID-S/RAID-5 performance to suck, even
with a fairly small cache. If you can write the whole stripe at once
there isn't any performance loss, your only "overhead" is calculating the
parity, which ought to be easy work for even cheap embedded CPUs these
days (let alone what EMC can afford to use in their flagship DMX)

Its when you don't need to write the whole stripe at once you are screwed
since you have to read the stripe and then write the stripe, but I think
if the algorithm was intelligent it should work with a nice big cache.
Just hold all those partially written stripes in cache and wait for the
time when you write enough parts of the stripe that you can write the
whole thing or when you read that stripe (or something near enough to it
that you might as well read it since the head is in the neighborhood)

I don't know whether the DMX does things this way, or have any direct
experience with its RAID performance, but consider that what they are
claiming is that the DMX using RAID is as fast as the 8000 series using
mirrored disk. NOT that it is as fast as the DMX series using mirrored
disk! It may be that they didn't change anything at all but made enough
improvements with the DMX that the statement is true. But it doesn't
necessarily mean that the ratio of performance between RAID and mirrored
in the DMX isn't the same as it was in previous models if it isn't out
they only changed the name of RAID-S, and not the implementation.

--
Douglas Siebert

"Suppose you were an idiot. And suppose you were a member of Congress.
But I repeat myself." -- Mark Twain