« Everyone Has a Tandy - The Human Mind and the Evolution of Information Technology | Main

July 21, 2010

Comments

What model was this run on, and what model controller?

Hi Don,

My IOPS testing was on a FAS3140 with OnTap 7.3. However, the model is pretty much irrelevant as it was more of a test on what a single drive can do.

The customer was using a FAS3160 with a 256GB Flash Cache card.

Disclaimer: NetApp beginner :)

Mike, regarding the random I/O reordering, does this mode of operation hit a ceiling at some aggregate utilization percentage? E.g. when it's getting harder to find full empty stripes? I think I saw that mentioned someplace but can't find the reference now. What would be the threshold to watch out for?

Further, I can see that the reordering also easily avoids read-modify-write cycles for RAID-DP's horizontal parity. What about the diagonal parity? Doesn't it require additional reads from surrounding blocks which would introduce random I/O again? Or is WAFL trying to gather enough stripes so that it can calculate the diagonal parity fully out of memory too?

Sebastian,

It seems like you are pretty much on track. The efficiency does change as there are less free blocks and fewer empty stripes to choose from. But even on nearly full systems, the performance never gets nearly as bad as on traditional fixed-block arrays because choosing 5 stripes on nearby tracks vs. 1 stripe on one track is still better than dozens of stripes all over the platter.

There is additional logic with data placement on the platter itself. If a stripe near the edge of the platter is available, it will get a higher priority for placement. On a new system, most I/O will go to the edge of the platter. If you are like most customers who never fully utilize the capacity on an array, your data will live on the faster side of the platter.

Our competitors frequently show the performance curve from a new system to a completely full system relationally, without showing the actual IOPs between their array and others. In reality, our arrays can perform 200-300% faster on an empty system, but may drop to only 200% faster as the system ages. All of our designs for customer implementations as well as our public benchmarks are based on aged systems.

As for RAID-DP, there can be a very slight performance impact of about 2% on writes due to a second write on the DP disk. You can get all the details you want in this white paper.
http://www.usenix.org/publications/library/proceedings/fast04/tech/corbett/corbett.pdf

Thanks for commenting!

The comments to this entry are closed.

TRUSTe CLICK TO VERIFY