Question to the Storage Advisors … (Alex) We recently bought a Chenbro case with 24 drive bays, there are 6 independent backplanes, 4 drives each. After a costly and dismal failure with “green” drives we bought some WD enterprise 2tb sata devices. To fuel this array we purchased two 51245’s, I know we could have gone with a single 52445 but I was trying to engineer the most disk IO I could and our motherboard has 7 8x pci slots (Asus p6t7 supercomputer motherboard).
My initial plan was to have 2xRAID6 arrays. Each card would host 1 array, I put 11 drives into each with a remaining hot spare. This would have left me with 18 drives worth of data for storage. Using LVM in Linux the plan was to stripe across these, so I guess its a 2-controller RAID60.
There was a huge problem though, after letting the array build and verify over a weekend the performance was awful. only writing at 200MBytes/sec (1 byte being 8 bits, just to be clear). After playing around for a day with various sized RAID0s and RAID5’s (skip init), to make sure the disks and back planes weren’t the issue, I decided that a 4x raid 5 would probably work pretty well. In this setup each card runs 2xRAID5s, 5 drives each with 2 hot spares, in total leaving me with 16 drives worth of storage. I let this new config build and verify over night.
The next day I did some tests stripi
Alex,
Unfortunately something went wrong with the blog and I did not get all the story, but I think I get the story (to to speak).
All of my questions would revolve around “what sort of data are you using here”. If this is sequential data then there is an issue. If this is random data then there is an issue of another kind. Random data does not sit well on parity arrays. The overhead of small writes to parity arrays (eg database) kills write performance.
Also … quick init/skip init should never be used unless talking to a RAID tech when you are sweating profusely and trying to get data back from a lost system. They put the array into “full stripe write” mode … on which write performance absolutely sucks.
So …
Please reply back to this with information on the kind of data you are running on this system and how you are testing this (ie what benchmark etc).
Ciao
Neil
- Share
-
-
-
-
-
-
Send to a friend
-
more...
- | Post a Comment
