Data center designers should take a page from the High Performance Computing (HPC) market and standardize on a homogeneous infrastructure, rather than mix and match their hardware from a variety of suppliers, argue some analysts.
Even though HPC and virtualization present very different usage scenarios at their core, the two share some similarities. Both utilize a pool of resources for a variety of jobs in the most efficient way—but how they get there is where the two differ.
“The essence of virtualization is about abstraction. It’s a separation of a logical concept from the underlying physical reality,” explained Peter ffoulkes, senior analyst with TheInfoPro (a division of 451 Research).
Virtualization is, essentially, putting multiple workloads on a single piece of hardware. Most virtualized workloads are actually bandwidth-intensive, stressing the storage and networking systems more than the CPUs. The difference is that virtualization is designed to run multiple loads on underutilized CPUs in order to increase their utilization; the resources of an HPC system, by contrast, are often maxed out, and need to be doled out to the researchers competing for them.
The only place they differ is in the software layer. In the HPC world, a hypervisor would just slow the system down, so they’re not used. Instead, HPC uses scheduling software that acts as a queuing system. People put in their requests for CPU time and the software looks at what’s available, carving up resources to meet business priorities and making sure jobs are run at the maximum efficient usage of the machine.
In the past, the conventional wisdom for filling out a data center centered on one idea: don’t put all your eggs in one vendor’s basket. As a result, data centers are often a mish-mash of IBM, Dell and HP, especially if they have been bought over time.
But in an HPC system, you buy a whole lot of one type of system and have hardware uniformity across the board. The Top500 list supports this: the systems are uniform, with the same vendor, product/part (rack or blade) and the same CPU.
That is something virtualized environments should copy, ffoulkes said, and for the same reason as HPC. “If you want to have something you know will behave in a consistent way, then knowing it’s running on an identical environment means you can test it once and know it will expect to behave in a certain way the same way on every machine,” he said. “If something goes wrong you got a lot more work to isolate the actual cause.”
Take Southwest Airlines, which has one airplane model: the Boeing 737. United and American are a mish-mash of different Boeing and Airbus planes. “Southwest has been successful because they standardized on a single aircraft design,” ffoulkes said. “They only train people once, they got one set of parts and if there’s a problem at an airport, they know they’ve got the parts there. Any airline with more variants of aircraft has a bigger problem. That’s why people want to standardize on hardware. You can get to a root cause of a problem faster.”
This strategy works well when building a whole new data center or replacing one en masse. For companies that can’t (or won’t) replace their systems all at once, ffoulkes still recommends the single image strategy—just do it in a slow, graduated process over time.
Andi Mann, vice president of Strategic Solutions at CA Technologies, said he agrees with ffoulkes to a point. “I think it’s a good idea, a best practice, to standardize on a hardware build and hypervisor,” he explained. “It reduces the fragility of the environment and gives you the opportunity to have stability. But there are a lot of real world circumstances where it’s not a good idea. The cost is one. Do you need the same system for every workload?”
In such a case, one size might not fit all. For example, paying $1,500 per VMware license for all environments might end up wasting money.
Zynga, for example, launches new games on Amazon’s public cloud until it knows how big the load will be and how many people will play it; then the company deploys the software in the appropriate environment. Some games need more resources than others, so in some cases, there might be too many or too little resources available.
It all comes down to addressing the concerns over vendor lock-in. A smaller company might end up buying everything from one vendor anyway, simply because its scale doesn’t justify multiple vendors.
ffoulkes doesn’t advise major global firms to adopt a single part strategy across the board for a variety of reasons. Instead, they should deliberately spread their risk across vendors, make them bid against each other, and divide up virtualization pools in different data centers among different vendors.
“The bottom line is you want to keep your virtual cluster the same,” he said. “In that sense, you’ve got a known set of resources and you know it works. So decide on the cluster size and have vendors bid on it.”
Mann, however, is more cautious. “Where you get into problems is refresh cycles,” he said. “If all of a sudden HP doesn’t make a certain model you’ve standardized on any more, you got a problem. Your standard systems atrophy. You cannot maintain a single system image.”
When building, updating or expanding a data center, ffoulkes said he advises the customer work closely with the hardware vendor to know their roadmap, so customers will how far out of spec they will be in 12-18 months of time and be able to accommodate that in the design. And that’s the bottom line: forge a relationship, don’t just sign a contract. It preserves both companies’ interests.