Customizable vs. Bespoke – Why HPC Cloud Struggles

It was the late 18th century, and man was harnessing the power of steam. The Industrial Revolution was underway, a movement that would forever change the face of the earth for both good and bad. In the middle of this cataclysmic change, an invention was born that would inspire Charles Babbage‘s Analytical Engine, the forerunner of modern day computers. This invention set in motion a chain of events leading to the Information Age, while changing the face of manufacturing forever.

This invention, first demonstrated in 1801, was the Jacquard loom.

Using a punch card system to program the loom, the weaver could instantly produce many different patterns. Though impressive as one of the first programming languages, the Jacquard loom did something else–something indicative of the entire Industrial Revolution. It ushered in mechanical production line processes. Faster, cheaper mechanical production meant the widespread use of handcraft to manufacture was at an end.

Interestingly enough, the Jacquard loom did offer something many of the other processes born in the steam-induced heat of the factory floor did not: customization. Regardless of what Henry Ford said regarding the color of his automobiles, customization had a way forward in this new world order.

But, is this a fair replacement for handcraft?

Consider the two following terms:

 
On the outset, they appear to mean almost exactly the same thing. But, there is a very important difference between the two. Namely, with customization one starts with an existing product and adapts or changes it. With bespoke, the product is created from scratch. In other words, with bespoke the product can be created with any given attribute, while customization only allows the modification of certain, predetermined attributes.

Cloud is customizable. HPC is bespoke.

The fact of the matter is when I go out to the cloud; I’m not really given

very many “low-level” decisions to make. In most cases, I’m only allowed to choose from some predetermined packages to which I may be able to make minor tweaks like determining the software stack. I likely won’t really know anything about the actual hardware—what CPU model, what memory types, bus speeds, interconnect, general networking, or another of myriad things.

And, really, that’s okay. I don’t want to care about those things. That’s not what I’m looking for from cloud. I’m optimizing for availability and ease-of-use.

HPC is different—especially at the high end. Here every hardware decision matters from top to bottom. Even the building where I store the machines needs to be engineered specifically for that purpose. Maximum performance requires me to care about memory configuration, networking hops, CPU parity, and the list goes on and on.

So, what do I do? I turn to a manufacturer to build me exactly what is needed. I match my workload to specific hardware specifications. I’m optimizing for utilization and speed.

The Bespoke Suit

Having a bespoke suit (or any other piece of clothing) made to order is a wonderful experience. Putting on the bespoke piece of clothing is… well… magical because it just fits. It feels good and moves perfectly in all of the right places. There is no stretching, straining or pulling. It was made specifically for you. It also has one ruinous effect—once one has experienced it, it is hard to go back to off-the-shelf clothing.

High-end HPC is no different. The supercomputer perfectly matched to its workload is unbeatable. It’s tuned so that each part works in perfect harmony with all the other parts. It is bespoke–perfectly created for its purpose. This is something cloud cannot yet offer.

It is true that some workload, like the types of jobs run by the BOINC network will work fine in a cloud environment. They are not so dependent on the hyper-tuning needed to squeeze every last possible compute cycle out of the processor cores. This can also be the case with many of the small- to mid-size clusters. If so, HPC in the cloud (or HPC Cloud as it is sometimes called) may be the way to go.

But, for the high-end in terms of size or throughput, the basic customizations available on the cloud and their inherent drawbacks aren’t currently able to provide what is needed. Perhaps in the future they will. But right now these HPC practitioners have tried on the “bespoke HPC suit,” and the fit is too perfect. Why would they want to go back to something off the shelf that is almost, but not entirely a good fit?

The clothes back in those days were made so much better than clothes are today. They actually took time to make clothes to fit a woman’s body. Today they make clothes that fit sizes, so it stretches to fit this and that.
~ Debbie Allen

While teaching at university, I frustrated countless students by giving the answer, “It depends.” Things are no different here.

Pick the tool right for the task. Just because you have a hammer, not everything is a nail. If you happen to fall into the magic middle ground where HPC cloud may work for you, consider tools like Adaptive Computing’s Moab that support you doing both.

Picture courtesy of Diego Torres Silvestre