IT Solutions Blog

Enterprise Class Storage Requirements from Mike Workman

In case you haven't read Mike Workman's (CEO of Pillar Data) blog,take a look at this latest post "Homey don't play dat."

Here is a snippet from the post where he argues the point of exactly what can be considered
"Enterprise Class" Storage: 

Let’s put down a few basic requirements of “Enterprise Class” that I would imagine we could all agree to:

1. I’d say at least 4-9s' of data availability, perhaps in many environments over 5. No single point of failure is necessary to achieve this requirement. Regardless of the failure, it should have limited consequences on performance or LUN access. Nobody claims a machine is Enterprise class when it is sold in single-controller models; This is just silly.
2. Systems should not be designed to assume there is “idle time” in every 24 hour window. Most Enterprises run 24x7, not 24x5, nor even 12x5.  It is clear when rebuilds, “data progression”, LUN layouts are stalled for days behind system load that they are not ready for the Enterprise.
3. Non-disruptive upgrades. This doesn’t usually include upgrades to new machines or new platforms, but code point-releases while system is in operation.
4. Failures during operation do not result in Filesystems or LUNs off –line by design. Not that they don’t ever go offline, but that he usually should not, and when they do it is a defect-not a design attribute.
5. Serviceability from the front or back of the rack.  In other words, field replaceable units do not involve pulling boxes out of racks and opening lids, using a pair of pliers or a soldering iron (see… hyperbole can be fun).
6. Reliability, which means that the quality of components in the system doesn’t necessitate constant intrusion in the data center. Regardless of #1, it is never desirable to have someone pushing a wheelbarrow of full of bad parts through the isles -- especially if it takes days to rebuild and restore the system to an all-normal condition, like in the case of many of our competitors.

A system which stops all I/O on a class of drives for 25 to 60 seconds when one drive fails is not Enterprise class. Yet there is a company out there in which this characteristic is indeed always present by design; a dumb design for the Enterprise to say the least.  The company is Compellent – or as I referred to them in my last post CML (their ticker symbol, this wasn’t a slight any more than people calling Sun “JAVA”).  Here are some of the test results for a Compellent system under I/O load. The figure below is a  plot of IOPs after a drive fails, versus the same thing on a similarly set-up Axiom. The IOPs were normalized to 100% for the no-fault performance on both arrays, and both had a similar number of spindles (both systems are current, not old stuff, and the code revision level of the CML system tested was 4.2.3 ).

 

Personally, while I recognize we all make mistakes, we all have flaws, and none of our systems are perfect, I believe that some of statements and claims in the storage industry are just deceptive, or reckless to say the least.

Read more about Pillar Data including the latests posts by Mike Workman by clicking here.