DL380G6 Raid 10 write performance

[Warren Beckett]

Hi,

We have several new DL380G6 with P410i Array controllers w/ 512Mb BBWC and 6 300GB 15K SAS drives running RHEL5. All of the latest firmwares have been applied ( 1.66 ). The servers all have 36G RAM, and both Sockets populated with quad cores.

I am wondering what is the expected sequential write performance on a single RAID 10 configuration, or in a RAID 5 configuration, both using all 6 disks.

What is puzzling to me is that for large sequential writes ( writing 100GB files ) RAID 5 writes twice as fast as RAID 10. RAID 5 has a write average of 250MB/s while RAID 10 comes in at 120MB/s.

Is that to be expected?

Kind Regards,

Warren.

[Kopp2]

Hi all,
I have the same question.
Why Raid10 is slower as Raid5 or Raid6 ???
Tested with 4 and 6 SAS 15k 146GB HDDs on DL580 with SmartArray P400 512MB BW cache . (see attached screenshot).
Only Raid0 with 3 or with 6 HDDs is faster as Raid5.
Someone an idea?
Regards,
Juergen Kopp

[Warren Beckett]

Hi,

I raised this HP support, and was asked in the end to upgrade the Firmware for the raid controller to version 2.00, which for us solved the problem.

Regards,

Warren.

[TTr]

I have done similar tests even with larger array systems and have found similar results. I actually found an optimal number of disks for my raid5 array groups. The number of disks depends on the individual array system so you should run these tests and decide for yourself and your array system. As you start with a small number of disks and then as you increase the number of disks, perfromance increases as well up to a certain number of disks (usually 6-8 disks). Then after that as you add more disks perfroamnce levels off and starts to decrease. That's to be expected, there are too many disks to collect from and process the raid stripe.
My explanation for this is that new array systems have much faster disks and faster and usually independent i/o buses than older array systems and the number of disks (iops) really makes a difference. More disks = more iops=more data. The controllers are faster too so they can process the data faster as well so in the same amount of time, a 6disk raid10 will generate almost half the amount of data as a 6disk raid5.

So during reading, a 6-disk raid10 really uses 3 disks for i/o whereas a 6-disk raid5 uses 5 disks for i/o. More reads in the same amount of time=more data faster. Writing is different because all disks have to be written to and raid5 still comes out atop.

You have to use large files in the testing so that you defeat the caching effect and you are really measuring disk performance and not cache performance.

It would be interesting to compare a 10-disk raid10 with a 6disk raid5. R10 might be better in reading but not in writting.

What is ironic is that even the vendors themselves are sticking to their conventional claims of raid10 being faster.

Sorry but I can't be more specific. Do your own testing and think about what the controller and disks are doing in each case.

[wobbe]

Interesting results but I do think that the tests performed are a bit limited. Raid five and also raid 6 usually perform poorly when it comes to a lot of small random writes. This is because of all the parity data the need to be calculated and written. You should also compare both systems under these kinds of work loads. But if you only have to do sequential writing of large files raid five or six could be interesting. I might look into this since Iâ ll be setting up a new server to store lots of hard disk images.