RAID5 full stripe writes on the XP12000

[Chad Sellars]

Everyone always says use RAID10 for logs. RAID10 has an efficiency of 50 %. 1 host IO results in 2 array IOs (1/2 = 50%) Database logs consist of a pure sequential write stream. Such a stream is highly amenable to full stripe write aggregation. (Open-V) During a full stripe write, there is no read and full spin penalty that is incurred by RAID5 during small block random writes. Consider a RAID5 7 + 1 full stripe write. 7 spindles receive data and 1 spindle, the parity, is overhead. This yields an efficiency of 87.5% (7/8). Naturally, cache write pending must be a healthy value. Am I wrong here?
Did the much maligned RAID5 just whup RAID10?

I created a test to see if this really happens in RAID5 7+1. Using iometer, I executed a series of random writes with block sizes ranging from 1K to 16M. I used performance advisor to monitor LDEV metrics. I was unable to see any of the reads that were occurring inside the raidgroup during 1K random writes. I had hoped to see the 2 reads for every host write. I had hoped to increase the write size up to the stripe size until the reads reduced. It seems that PA only provides metrics for the raidgroup or the LDEV and not the individual spindles. Is there a way to get spindle metrics in PA? Has anyone been able to see the reads that go on when executing pure random small block writes?

Since I cannot complete this test on the XP, I will conduct the test on the EVA with EVAperf and TLVIZ.

[Nigel Poulton]

Chad,

You're right about the RAID5 whipping the RAID10. You will also find that some vendors do actually recommend RAID5 for predominantly sequential write workloads such as database logs. But they almost always have a hard time over this. I recently did some work for a client where we recommended RAID5 for Exchange logs but MS and Avenade recommended RAID10.

As for being able to see spindle related metrics - Ive not seen this. Although it will now start to bug me as well ;-)

Regards

[Alzhy]

Hi Chad,

I am one of those students of thought that on the XP12000 (and similar Hitachi based architectures) Oracle Storage should be on RAID10 LDEVs/ArrayGroups. We've traditionaly used an ALL RAID5 (Open-V) approach (7+1P and 6+2P) for "everything". We usually have a standard 50-100 GB standard sized LDEVs. On the host, we do additional Striping of these LDEVs -- 4-8 ways and pathed on 4-8 FC CHannels. But still - there seems to be issues with Oracle always complaining of poor I/O or very lengthy I/O waits. On the host level metrics though - I see no evidence. No disk queueing (which I've wholly relied upon as the primary indicator of a slow LUN). DBAs are saying they've optimally done all they can do. We're using Cooked Filesystems -- DirectIO (VxFS).

So my final suggestion is - let's move from RAID5 LDEVs to RAID10 configs. Most of colleagues I know have never been adventurous in using RAID5 LUNs on their EMC, EVAs, Hitachis, etc... They always use RAID10 or any non-RAID5 layouts. The bean couneters and skeptics are asking though -- what are the expected gains if any? I simply answer them - the greatest benefit I see with RAID10 is better scaleability and better performance during crunch time situations. I could not however give them balpark figures. I did suggest getting tests done -- since we can easily clone to RAID10 sets.

So -- what do you think? Is switching to RAID10 really worth it? Our Applications behaviour is "anything goes".. It is supposedly OLTP but we'ev a number of "online" apps that behave like batch/DSS continously lurking in the backgroud doing continous updates and heay writes at times...

Any insights?