RAID-DP

[raadek]

TTr,

I am not saying there is no single vendor out there being able to overcome (or minimise) write penalty issues for RAID5/6 (NetApp being one of the examples).

Yet in the HP world we are looking at either SmartArray controller, MSA or EVA. In all these cases short writes on RAID5/6 get a significant performance hit due to underlying mechanism well explained above by Patrick.

Even on mighty EVA the best practice published by HP folks says "use VRAID1 for performance".

I've done a number of performance tests on traditional arrays comparing RAID10 with RAID5 (same number of spindles) & the former one just trashes the latter.

It's just the maths - to write a single block on a (traditional) RAID5 set, you have to read the parity info, calculate new parity, write parity & eventually write the actual data.

Rgds.

[Joshua Small_2]

Whilst RAID1+0 can be demonstratably faster in a test environment than RAID5, it's dangerous to take this at face value.

I've seen one group that purchased domestic SATA drives for a server, just so they could afford to use RAID1+0, because they read somewhere that it was "better", and noone could consider that the comparably priced RAID5 solution involving 10k SAS disks would be better.

I've worked with a number of slowly performing Exchange servers and it usually comes down to other services on the machine. Groupshield on your Exchange server has a terrible performance impact for example.

And if your deployment is big enough, Microsoft will suggest you deploy your server roles on different servers. Feeling you can avoid that be using RAID1+0 will definitely be wrong.

Only after you've considered the above, and the likelihood of networking (search these forums for performance issues with Hp server NICs at the moment), memory or similar issues should you even concern yourself with the whole RAID1+0 vs RAID5 arguement.

[Patrick Terlisten]

Hello,

don't forget that NetApp is using a simple trick: They cache write IOs in a NVRAM, so that they only write complete RAID stripes. A short write is often handled completely in the NVRAM. For security NetApp writes the NVRAM periodical to disks. They eleminate the write penalty simple by using a NVRAM.

This method could easily added by HP, but they don't do it. This would speed up every RAID level. But IMHO it's unsecure to do it all in a NVRAM.

btw: That RAID-DP is slow, while it's based on RAID 4 is one of the "silver bullets" which is used for marketing against NetApp. ;)

To make this story short: Use RAID-DP if you have a Filer. Use RAID 1+0 oder VRAID 5 if you have an MSA or EVA.

Best regards,
Patrick

[Riley Martin]

Joshua,

I'm new to Exchange. Could you explain further on Microsoft wanting server roles on different servers in a big deployment. Which roles, how big? Thanks.

[raadek]

Riley,

In Exchange 2007 you can have the whole bunch of roles: http://technet.microsoft.com/en-us/library/bb124935(EXCHG.80).aspx

In 2003 it is just about front-end/back-end split: http://technet.microsoft.com/en-us/library/aa996980(EXCHG.65).aspx

Patrick - you are spot on with the risk associated with NetApp NVRAM caching; that's why they mirror its content in the active-active config (it means using only half of the memory for actual caching).

Rgds.

[Joshua Small_2]

Being new, the real question to ask is, how big is your deployment?

If you have 20 users for example, I would suggest most of this thread is redundant, and complicating things by involving RAID levels will jsut confuse the issue.

If you have 10,000 users, there is far more to consider than what has been raised.