As the read/write ratio moves closer to 1:1, the RAID 5 write penalty of 4 becomes a much larger factor in determining the spindle count required for adequate performance. Exchange 2007 with cached mode clients in particular has a read/write ratio very close to 1:1. For the same number and type of spindles, you'll get twice the performance out of RAID 10 as you would RAID 5 in such an application workload environment. The rule of thumb is: If the read/write ratio of your application workload is less than the write penalty of your proposed RAID type, then your proposed RAID type is poorly suited for your application workload. You'll simply waste spindles.

All that said, on the extremely low end of the small business environment, say 10 Exchange users @ 0.3 IOPS/user, you're right; even a poorly suited RAID choice won't make any difference. The IO load is too small to matter. When you start to scale, that's when it gets ugly. Most small business owners I know have a dream, and for the most part a fairly solid plan, of expanding their business… Let''s not paint an overly rosy picture and end up putting a stumbling block in their way.

John

As the read/write ratio moves closer to 1:1, the RAID 5 write penalty of 4 becomes a much larger factor in determining the spindle count required for adequate performance. Exchange 2007 with cached mode clients in particular has a read/write ratio very close to 1:1. For the same number and type of spindles, you'll get twice the performance out of RAID 10 as you would RAID 5 in such an application workload environment. The rule of thumb is: If the read/write ratio of your application workload is less than the write penalty of your proposed RAID type, then your proposed RAID type is poorly suited for your application workload. You'll simply waste spindles.

All that said, on the extremely low end of the small business environment, say 10 Exchange users @ 0.3 IOPS/user, you're right; even a poorly suited RAID choice won't make any difference. The IO load is too small to matter. When you start to scale, that's when it gets ugly. Most small business owners I know have a dream, and for the most part a fairly solid plan, of expanding their business… Let''s not paint an overly rosy picture and end up putting a stumbling block in their way.

John

My point is that lots of systems use 4-drive SATA RAID-5, and lots of these are in production/corporate environments, and lots don't have crazy cache or processors. It's not ideal, and not even high-performance, but it works. And the Iomega ought to work, too, in an appropriately low-demand environment. Even with Exchange or SQL Server or ESX. I certainly don't expect 1-second IO waits. You just can't tar something as unacceptable just because it uses RAID-5!

As for the disks, I still contend that modern SATA drives (of the kind to be used in the Iomega and similar devices) can handle much much more IOPS than you're suggesting. Yes it'll vary greatly by IO size, and yes it might dip to the 20s in some situations. But modern drives can easily average 80 or 90 IOPS with real-world business application loads. Look at Tom's Hardware tests and see for yourself!

Most small offices use cheap bare drives, perhaps software RAID, and fast (not gigabit) Ethernet. I have no doubt that one of these low-end RAID systems would be a major major improvement for them! I have no doubt that an ix4 with iSCSI would be faster in all circumstances than an internal bare drive or two.

Thanks so much for the stimulating conversation. I really respect your thoughts, and hope you will continue to call me out and provoke discussion!

Where P is the performance in IOPS/spindle and N is the number of spindles in the array:

RAID 5 random write performance = P*(N-1)/4
RAID 10 random write performance = P*N/2
RAID 5 random read performance = P*(N-1)
RAID 10 random read performance = P*N

Yes, SATA drives have a MAXIMUM IOPS that is higher. The point is that an IOPS value is meaningless unless ascocated with an IO size and a response time. Sure, I can push a maximum of around 90 IOPS per SATA spindle at an IO response time in excess of 1 second. If each outstanding IO took a second to complete, my computer would appear frozen. A practical response time for random IO response time is 20ms for many applications and for databases in particular, closer to 10ms. I stand by those numbers.

In the product release, you'll note they gave an example of an application that uses large sequential reads… meadia straming. Not pounding on Iomega, but this is standard vendor fare to put their product in the best artifically crafted workload light.

Still, this is targeted clearly for the SOHO market. How much random IO can a thread or two produce? Are home standards as exacting as enterprise standards? I think not. Add some cheap replication and a second unit and even I could see the potential for home office use. The price seems right. I wouldn't rush out and attach one to the SQL or Oracle server backending my revenue generating Web application,or even Exchange server for that matter, (if I had one-not specifying that I do or don't).

John

I see four slots on the front. 1TB or 500G B drives to get 4TB or 2TB RAW then. lET'S SEE, 38 random IOPS per spindle @ 20ms response and a RAID 10 wite penalty of 2 …ouch. These are definitely not enterprise class. You'll hit a disk bottleneck long before iSCSI factors become a consideration in the performance calculus.

Still, this is targeted at SOHO. Buy two and some cheap host based replication software…

“Serious iSCSI performance” – if I wanted that, I wouldn't be buying sub-$1k gear.

<sigh> Dang it, so much for my free lunch.