RAID Cards performance issue

[Christopher Chan]

Scott Balneaves wrote:
> On Mon, Dec 08, 2008 at 08:21:38AM +0800, Christopher Chan wrote:
> 
>> Software raid5 get performance penalties but a hardware raid card with 
>> sufficient cache memory (must be battery backed if you want to minimize 
>> data loss) and processing power can do raid5 and perform as well as or 
>> even better than raid10 depending on the number of drives involved.
> 
> No, RAID5's a compromise.  If it's a compromise in software on the server,
> it'll be just as much of a compromise on a dedicated controller, where
> it'll be implemented in software running on the card's controller.

When I say 'Software raid5 get performance penalties', that is in 
relation to the performance you can get with hardware raid cards. That 
is, in this day and age.

Popular hardware raid cards using the puny and useless Intel i960 
processor a decade ago were absolutely creamed by software raid5 in 
performance.

Even if you do have the processing power on the hardware raid card, you 
still need sufficient buffering for the processor as the tests done by 
the Gleb research group in the link below will tell you.

http://www.chemistry.wustl.edu/~gelb/castle_raid.html

The 3ware 850x series has no cache memory. They will therefore perform 
absolutely poorly in raid5 mode. As the tests indicate, hardware raid5 
performance is really poor on the 3ware 8506. However, when they 
switched the card into JBOD mode and used Linux software raid5 to 
implement raid5, the performance they got was comparable to the 3ware 
hardware raid10 figures. If that is not an argument for raid5 
performance, I don't know what is.

Today, bus traffic on the mainboard gives software raid5 performance 
penalties compared to a hardware raid card doing raid5 since the disks 
are directly connected to the raid board/processor. But the raid board 
needs sufficient processing power and onboard cache to be able to pull 
off the performance.

> I've run tests myself on 3Ware controllers, and they are MUCH slower in
> a RAID5 config than either a RAID1 or RAID10 config.  This isn't a smack
> at 3ware controllers: I use them myself, and they're great.  Solid, dependable
> raid controllers.  It's just a limitation of RAID5.  RAID5 tries to be
> everything to everybody ("More Space!!" "Fault Tolerant!!" "Less Filling!!")
> and in the end, doesn't really satisfy anyone.

Your problem is most probably because you did tests with the 3ware 750x 
or 850x series. I know those boards suck at raid5 from personal 
experience. I also know that newer 3ware boards with cache memory have 
solved the raid5 performance disparity. I had a ten disk raid5 array on 
a 3ware 9550 that was loaned to me and I used it as a mail queue and it 
rocked. I was the MTA guy at Outblaze Ltd. at that time.

Here is a newer test that also uses six disks like the Gleb research 
group did but with a different controller that has both the processing 
power and sufficient cache memory.

http://www.linux.com/feature/140734

Not surprisingly, raid5 beat the pants off raid10. Why? For six disks, 
on raid5, you have all six disks for input and output. On raid10, you 
need to make three mirrors and so you are effectively reduced to three 
disks for input and output. Given sufficient resources on the raid 
board, raid5 with six spindles will beat raid0 with 3 spindles. Raid5 is 
effectively raid0 + uber processing. Therefore, if 'uber processing' 
keeps up, you are really doing an unfair raid0 with 6 logical disks 
versus raid0 with 3 logical disks knockout.

You also get software raid performance comparison in that article and 
you will see what I have said about software raid performance suffering 
performance penalties proven except for one thing that I found most 
interesting...software raid 5 got the best read performance and that 
simply blows my mind away.

Enjoy.