Estimating Storage System Capabilities Should not be a Risky Business!

Blast from the Past

Remember the 1980s? I am sure most of the people reading this were around back then. It was a time of big hair, boom boxes, and valley girls. Kids played Nintendo, watched the Smurfs and tried to solve Rubik’s Cube. Every little girl wanted a Cabbage Patch Kid. On television, commercials asked us “Where’s the Beef” and Max Headroom became the world’s first computer-generated TV host. It is normal to be nostalgic about the past. Life seemed simpler then and everything tends to look rosy through the lens of history

Roots

In many ways the 1980s were the dawn of our current technological age. The personal computer entered common usage in business and at home; mobile phones became more than just a novelty, and the internet was born (and Al Gore was not the father!). Getting more specific, the first cached Disk Storage Controller, the IBM 3880, was introduced in 1981. Cache was one of those fundamental changes that revolutionized storage. It also required a paradigm shift when looking at disk storage performance. Yet some still cling to the past in this area.

Max Headroom

Although our friend Max is virtual, the concept of headroom is quite real and can be very useful. It tells you how much more growth a resource can sustain before you will have problems. I have seen headroom reports from several infrastructure monitoring products that purport to tell you how many IOPS you are running compared to the max IOPS supported or “ceiling”. The ceiling is calculated based on the number of disk drives you have installed. Although this may have made sense in the 1980s, the advent of disk cache and sequential prefetching made this concept obsolete.

Ghostbusters

If you want a useful headroom metric you need to define it properly. For example, using a max IOPS ceiling based on front-end I/Os does not make any sense. Lots of front-end I/Os may be cache hits and never actually use the back-end drives. Sequential read patterns are detected by the controller generating a pre-fetch so that data is streamed from the disk with no seek time. Write I/Os are always cached. Although writes must be eventually de-staged, modern storage controllers will gather writes in an efficient manner so that the total backend load will be less than if written directly to the storage. Finally, RAID overhead may generate additional backend operations not included in the front-end I/O rate.

Back to the Future

There are several more sophisticated views of headroom that are applicable to modern storage systems. One is to base the ceiling on throughput. Throughput is much more representative of the work a storage system is performing and reasonable thresholds may be estimated if you understand the hardware’s architectural capabilities. Another option is to base the ceiling on back-end IOPS. If you have workload information such as cache hits, read and write %, sequential % and RAID configuration it is possible to calculate the number of random and sequential back end ops per drive. The ceiling for this metric would be based on the number of drives in use and published drive specs.

Top Gun

If you want to rule the data center the way that Maverick ruled the sky, you need both skills and the right instrumentation. IntelliMagic Vision provides you with the intelligence to make timely and accurate decisions about your IT infrastructure. If you would like to see how IntelliMagic Vision can help you, please email sales@intellimagic.com to request a demo.