If I had an F230 with, say, 6 drives, would I see any significant performance gain by placing 3 drives per shelf, rather than having all 6 drives on one shelf (and thus one controller)? This will be for heavily randomized file accesses (e.g., mail or news spool).
I would be willing to bet that this is probably one of those questions only answerable by testing for accurate results.
Seems like the fewer drives/controller should be faster, but whether or not it's measurable would be an intersting intellectual excercise.
Since the toaster only uses 1 parity disk for a whole pool of drives, I'm going to bet it's not significantly faster.
On Wed, 24 Sep 1997, Brian Tao wrote:
If I had an F230 with, say, 6 drives, would I see any significantperformance gain by placing 3 drives per shelf, rather than having all 6 drives on one shelf (and thus one controller)? This will be for heavily randomized file accesses (e.g., mail or news spool). -- Brian Tao (BT300, taob@netcom.ca) "Though this be madness, yet there is method in't"
On Wed, 24 Sep 1997, Jaye Mathisen wrote:
Since the toaster only uses 1 parity disk for a whole pool of drives, I'm going to bet it's not significantly faster.
For writing perhaps, but disk reads don't touch the parity drive. I'm going on the assumption that two controllers can queue up more commands than one. I suppose having a lot of read/write cache will mask most of the benefit of splitting drives across controllers though.
If you don't run noatime, you'll generate significant writes, depending on access type as I recall, and there's other things going on...
I'm still going to run with "not significantly faster". :)
On Wed, 24 Sep 1997, Brian Tao wrote:
On Wed, 24 Sep 1997, Jaye Mathisen wrote:
Since the toaster only uses 1 parity disk for a whole pool of drives, I'm going to bet it's not significantly faster.
For writing perhaps, but disk reads don't touch the parity drive.I'm going on the assumption that two controllers can queue up more commands than one. I suppose having a lot of read/write cache will mask most of the benefit of splitting drives across controllers though. -- Brian Tao (BT300, taob@netcom.ca) "Though this be madness, yet there is method in't"
Jaye Mathisen wrote:
If you don't run noatime, you'll generate significant writes, depending on access type as I recall, and there's other things going on...
I'm still going to run with "not significantly faster". :)
I would tend to agree, given the original premise of random I/O, since most of the time the controller will just be waiting for seek completes, no matter how deep the tag queue is. Having more controllers should help if there is significant sequential I/O, since you should be able to get multiple parallel streams going. With WAFL, this should happen a lot for writes, so it is probably a modest win overall.
Mark Muhlestein -- mmm@netapp.com
-
Brian Tao -
Jaye Mathisen -
Mark Muhlestein