Is it normal for Windows Software Raid (Mirroring) to slow a drive's performance?

Question

I've been doing some performance tests on a few of my drives lately, trying to figure out which combination of drive/raid-level(or not) will give me the best performance (for Hyper-V incase anyone's curious).

However, I'm seeing some behaviour/results that seems quite odd to me.

I'm using standard Windows Software Raid (ie, not fake motherboard raid, and don't want to spend the few hundreds of dollars I'd need to in order to get a decent hardware raid controller) for these tests. I'm not at all concerned about backups or drive failure (Windows Home Server, FTW!), I'm focussing strictly on performance. These tests are being executed in Windows 7 Ultimate (My server will end up being Server 2k8R2).

I've got two Western Digital Black 640Gb drives (SATA300) which give the following results using Atto Disk Benchmark (I also get similar results with SQLIO):

Drive               Read    Write
640(#1)              120      115
640(#2)              125      110
Mirror (Raid-1)*      82       77
Striped(Raid-0)      162      145

The unexpected result is the drastically poor performance fo the mirrored drives. I was expecting reads to be slightly better than a single drive, not significantly worse..

I've run the tests a number of times (breaking and recreating the array and mixing up the mirror vs striped vs single drive tests) with comparable results.

For anyone using a similar software raid setup, do these numbers jive with what you're working with, or are they completely off base? (Because the certainly seem so to me...)

Solution 1

If you don't wait for the array to sync after creating the mirror, you will see results like this. The drives need to be in sync to have a real measure of performance taken, otherwise they will be doing a lot of background reads and writes.

Solution 2

Long time ago, but responding for others ending up here from Google.

First, synching does so on as sector by sector basis, whole drive, data or empty, partitioned into 1 or many.

Second, during synching, for some reason I have not been able to fathom (yet), the speed of the drives drops down to a fraction of normal speed - maybe some buffering effect, caching, I don't know.

When the mirror set is not synching, then you should find that the mirror set is faster than a solitary drive. This has to do with intelligent use of parallel disk accessing. My data drives are 7500rpm and HDTune shows them as fast together as my 10000rpm velociraptor.

Comments

[Peter Bernier]

If I had just created the array, wouldn't it be blank and without any data to sync?

[MDMarra]

@peterbernier - Not necessarily. You can add a mirror to a drive that already has data