RAID10 writes faster than reads?

J

joblo37pam

2[H]4U
Joined
Jun 28, 2002
Messages
2,211
I'm playing around with a new storage system for my in-house esxi host (running SBS11, untangle, and a couple VMs for testing) and getting some strange results that I can't explain and was hoping the experts could help me out.

I have a PERC6i controller and 4 Samsung 2tb F4 drives. I am trying to do some experimenting with RAID5 vs RAID10 before I make a decision of which I would use permanently, but the results aren't making much sense to me.

First off, I updated the firmware on the controller to the newest that Dell has released. Then I created the four drive RAID10 array with a default 64KB stripe size to start testing. I installed Win7 Pro just as a basis for testing, and ran CDM. I didn't take a screenshot (I can rerun and post one if needed), but I got a max read of 146MB/s and max write of 256MB/s. I'm not sure what could be causing the reads to be slower than the writes and don't want to do a ton more testing or spend the time to let a RAID5 array build before I figure out what's going on.

I previously had this controller running a 3 drive RAID5 array, and although it has been quite a while since I had run benchmarks, I don't remember anything strange like this.

Does anyone have an explanation that might explain this odd behavior? Any help would be appreciated. Thanks.
 
Although the read speed is low I suspect cache on the raid card is causing the higher write speed. Do you have write back cache enabled?
 
Cache is affecting the write speed. As for the read speed, it's possible that the Perc6i is only reading from one drive in each mirror set. I believe this is a limitation of all software RAID and some entry level hardware based controllers. It would make sense with the numbers your seeing.
 
I believe this is a limitation of all software RAID
Click to expand...

linux software raid will read from more than 1 disk at a time in raid 10 or at least that is what I am told. I do not use raid 10 very often since I generally put 8+ drives in a server and the bulk of my data is medical images.
 
Last edited:
Would the cache still have that big of an effect with the bigger test sizes in CDM?

I also noticed that the array is doing a 'Back Init'. Status says it is normal, but when I originally built the array I chose the fast initialization, so I assume the 'Back Init' is just finishing that process. I'm guessing that may have something to do with the performance, or is 'Back Init' run at a lower priority so it doesn't affect throughput?
 
Ok, I let the initialization finish overnight and the reads improved a bit, but it is still a a little slower than I would expect, so I ran a few other benchmarks to try to find some explanation:

RAID10_CDM.jpg


RAID10_ATTO.jpg


RAID10_HDTUNE_READ.jpg


Those numbers seem to be all over the place to me, but I've never run these benches on a RAID array, so maybe they aren't abnormal. Anyone have some input?

I may just try to build the RAID 5 array and see what that looks like.
 
Ok, it just got a little stranger. I rebuilt the array as RAID5, again with a 64kb strip size, and this is what I got:

RAID5_CDM.jpg


I don't really know what to try next. Shouldn't the sequential read speeds of a 4 drive RAID5 array be ~30% faster than RAID 10, but with slower writes?
 
Did it finish initializing?

Sequential read speed should theoretically be the same as raid 0 with 3 drives or nearly 3 times the speed of a single drive.
 
It did finish initializing. I let it finish before I even exited the controller config, then restarted and installed Win7 again and the same drivers as last time.

RAID10 should be the same sequential speed as RAID0 on two drives, so if RAID5 should be sequential speed of three RAID0 drives, it should actually be 50% faster at reads, but this setup is showing the same read speeds in RAID10 and RAID5. Something isn't right.

Is CDM what I should be using to effectively test RAID speeds?
 
Ok, I ran ATTO again just to compare, and the numbers actually look better there:

RAID5_ATTO.jpg


So which one do I believe (or neither)?
 
You must log in or register to reply here.
Back
Top