Capacity – Failing to Plan = Planning to Fail

I am always intrigued by the notion that Cloud Computing should provide the “illusion of infinite capacity.” Illusion seems to belong to the domain of magicians. How does it fit into the realm of so scientific a discipline as Information Technology?

Most people are familiar with the quote from SF author Arthur C. Clarke (Clarkes’s 3rd Law.) “Any sufficiently advanced technology is indistinguishable from magic. “ All magicians use “gimmicks” and in this case, the gimmick is called Virtualization. Who would have imagined a decade ago that we would be capable of running 80 VMs on a physical server or bouncing them back and forth among hosts without user intervention? Those of us who struggled shoehorning multiple applications into a 640K memory space never dreamed of the capability of sharing 640 GIGABYTES among server Operating systems, “stealing from Peter to pay Paul” and not seeing these machines blue screen on a regular basis.

Capacity Planning in the old days meant considering the worst case scenario for our computing needs, and then purchasing MORE resources than were required (just in case.) The science/art even became codified under ITIL where an entire discipline is devoted to this practice. ITIL 3 (which is more oriented toward the business view of IT) views capacity management as comprising three sub-processes: business capacity management, service capacity management, and component capacity management (known as resource capacity management in ITIL version 2). All are inter-related.
So, let’s get back to the “illusion.” In virtualization, and especially in Cloud Computing, the key to providing this illusion is called “over-subscription.”

The practice of over-subscription predates virtualization. The Telecomm Industry and Disaster Recovery companies have relied on this model forever. Whenever resources are shared by multiple parties, there is usually the assumption that 100% of the users will never request the resources concurrently. Anyone who has ever shown up at an airport holding a boarding pass yet has still been denied a flight understands the potential fallacy in this thinking. HOWEVER, because the SLA to which the airlines are required to adhere typically states “in small print” that they will do their best to get you on the flight, you may still need to foot your own hotel bill, and pay for an extra day of parking until they have a resource available to accommodate you. If you are lucky, you might get a voucher for an atrocious airport meal.

Cloud Computing providers, at this nascent stage of the industry, are likely to offer SLAs about as flexible as those provided by the airlines. Just as you have the choice to seek alternative transportation, one can choose to continue “business as usual” and at best, adopt “cloud-like” practices for your own data centers. However, if among those practices, you are expected to provide the same “illusion” with your Private Cloud, then the onus is on YOU to determine how to over-subscribe safely and not run out of resources, and thus fail to meet YOUR internal SLAs.

Although it is hard not to blame the airlines (or the Cloud Providers) the fact is that it may be en economic necessity in order to function properly. Airlines and hotels hate to lose money on empty seats and rooms. Talk about perishable goods! Those are dollars that disappear and can never be recovered. And so it is with the Provider, who needs to buy the infrastructure to host VMs, pay administrators, utilities, Real Estate, etc. All the cost Cloud Users are looking to avoid are precisely the costs the Providers need to absorb.

Since the early days of virtualization, there have been tools available to aid in Capacity Planning. One we used frequently came from a company called “Asset Optimization Group.” Every time this tool was run, users saw how inefficiently their resources were being utilized; their immediate response was to sign a P.O. for virtualization software (which at the time meant VMware.) VMware purchased the entire company, and the product is now called “Capacity Planner.” Subsequently, many third-party CP tools arrived on the market. PlateSpin PowerRecon (purchased by Novell), NetIQ, Hyperformix, CiRBA Data Center Intelligence are all well-known offerings. Most vendors of hypervisors (Citrix, Microsoft, RedHat, etc.) will be happy to foot the bill for the rental of one of these as a Pre-Sales offering if your deal is sufficiently large to warrant it.

But once the deal is done, Capacity Planning doesn’t end. Before your virtualization sales rep has even cashed his commission check, it is likely you will be planning for expansion.

A gentleman named Keith Hanna from a company called TeamQuest forwarded me a couple of interesting white papers the other day. His organization is steeped in the ITIL tradition, and understands both sides of the CC fence (Providers and Consumers.) I do not know if they are available yet on their website, but they do have quite a few more “generic” papers on the practice of Capacity Planning and Performance Management (a closely related discipline.)

In a previous blog, “Cloud from both sides” I considered the very different perspectives of Providers and Consumers of Cloud services. Capacity planning is essential for both sides. Providers need to be profitable based on the resources they are renting out, and Consumers need assurance that they will not find themselves “on an overbooked flight” when they suddenly need to spin up a few hundred “unplanned” servers.

In terms of Performance Management, it might behoove a User to monitor their provided VMs to ensure they get what they expect. In multi-tenant environment, the virtual cores and NICS supporting your apps could be yielding to the increased demands of those of your competitors if a faceless and ruthless scheduling algorithm is doling out resources.

The “illusion” of infinite capacity is one of the most appealing aspects of CC. Just don’t let yourself be fooled.