:RAID 0. I know, I know, it sounds almost suicidal but as I :said speed is more :important in this instance than recovery. We'll be testing :this 760 against :an Auspex, which can be configured to use RAID 0, so I was :wondering if the :NetApps could also do this. ---
Actually, the way RAID4 is used in conjunction with WAFL, really looks and feels like RAID0, just with a parity write at the end.
WAFL removes the problem on a UFS type filesystem of having that single parity drive just thrash to tiny bits, taking the stripe performance with it.
When a series of writes comes into the box, it is buffered *im making this simple*, and it flushed out to disk (on a 760) ever 10 seconds, of when 16Mb of data is in the buffer (NVRAM).
A single stripe (or series of stripes) is then written in a _very_ optimized manner using the WAFL filesystem, and a single parity calculation/write is created per stripe.
I direct you to a document Mr. Hitz wrote at http://www.netapp.com/tech_library/3001.html for more details.
I have every confidence you will enjoy our performance in your testing.
As Jeff said, Netapp's RAID-4 implementation is not performance limiting in most cases. Most of the time a filer with RAID-4 will provide better performance than an equivalent product without RAID or with RAID-0. So I'm sure you will be quite pleased in your tests, particularly against a (*snicker*) Auspex. (Be sure to compare the price and number of filesystems in each system configuration as well.)
I seem to recall some public talk waay back when about Netapp possibly offering RAID-0 since basically all you have to do is remove the parity drive. You could simulate this yourself by manually failing the parity drive and letting the filer operate in degraded mode with just the data disks. However, I'm not sure that the code is actually optimized for this case, so there might be some other normal RAID-4 checks that get in the way of realizing any performance gains.
Bruce
On Wed, 29 Dec 1999, Mohler, Jeff wrote:
When a series of writes comes into the box, it is buffered *im making this simple*, and it flushed out to disk (on a 760) ever 10 seconds, of when 16Mb of data is in the buffer (NVRAM).
This is inaccurate, I believe in clustered machines with 32MB NVRAM it is 8MB. I think the general rule is half of NVRAM dedicated to the particular filer.
Tom
-
Bruce Sterling Woodcock -
Mohler, Jeff -
tkaczma＠gryf.net