February 3rd, 2005
A further look into utility computing and transparent pricing
After a few days of immersion in Sun’s orbit, you have to take a step back to access whether the well-crafted presentations about grids , Java, Solaris 10 , open source , subscription pricing and processor threads are the road to riches or a black hole in the making.
I admit I am a fan of the grid infrastructure and software-as-a-service models, and the notion of commodity-like pricing (not that the underlying technology is a commodity) is on the horizon. Sun has set $1 per CPU per hour as its “transparent” price for now, but as I wrote earlier this week, an industry standard definition of CPU per hour usage doesn’t exist. There is no equivalent to kilowatt hours or the price of a barrel of oil for CPU usage. It’s some combination of hardware, Solaris 10, memory, a grid engine, storage, and whatever else goes into enabling grid nodes to be used with minimal hassle. Provisioning and de-provisioning the servers has to be mostly automated or utilization rates will suffer.
Given the lack of a standard unit, Sun’s strategy is to fill the vacuum, get in front of the grid utility service trend and gather enough customers and momentum to create a de facto standard definition for CPU and storage metered pricing. As part of the push to define a transparent pricing model and compel other vendors of grid server farms to follow along, Sun is developing an online trading system with Archipelago Holdings (an online stock market) that will allow traders (corporate customers) to bid for computing power, using Sun’s definition of cost per CPU per hour as the trading unit. Robert Youngjohns, Sun’s executive vice president of finance and strategy, also envisions a futures market in which Wall Street companies and insurers might want to forward purchase CPU capacity for risk analysis in anticipation of a heavy hurricane season. HP Labs has devised a similar concept with Tycoon, a distributed, market-based compute resource allocation system.
I went back to David Gelardi, vice president of Deep Capacity Computing On Demand at IBM, to get a better idea of how IBM’s 50 cents per processor per hour on average compares to Sun’s $1 pricing. It’s definitely not an apples-to-apples comparison. The pricing is not transparent, which would make it impossible to have both as part of the same commodity (cost per CPU per hour) trading market.
model for its compute grid is to avoid customization. However, Youngjohns told me that if a customer has a big enough appetite for CPU cycles to justify tweaking the software stack, such as substituting Linux for Solaris, and pays for the license, Sun wouldn’t turn down the business. At this stage for Sun, getting the business is more important than the no customization credo.
IBM’s model is that every job is custom, and as a result, pricing is variable. “Each client will negotiate to build a software stack that makes sense for the problem they have to solve. Every partner has a different set of requirements–certain memory, stack, storage, and CPU, and we bake it into pricing model,” Gelardi said. Gelardi disputes the claim made by Jonathan Schwartz, Sun’s president, COO and chief transparent pricing evangelist, during his presentation at Sun’s product announcement that IBM’s price is closer to $5 than 50 cents and opaque. "Fifty cents is the midpoint of what we have sold to date. The price depends on the size of the stack, but we haven’t sold anything above $1 per CPU per hour," Gelardi told me. However, IBM charges a $5,000 annual "home base" fee (5,000 hours of CPU time in Sun’s world) for set up and storage of images. Gelardi also said that provisioning and de-provisioning is still largely a manual process.
IBM is using some older Xeon processors, which would run jobs more slowly per CPU than the newer Sun Grid Opteron systems, which are stripped down, with no front bezel, disk drives or PCI connectors.
In a recent blog posting, Schwartz invited IBM CEO Sam Palmisano to "fill in the blanks," in a chart comparing the elements (processor, memory, operating system, etc.) Sun’s grid utility pricing to IBM’s. No answer so far, and IBM’s business model is about variable pricing.
If customers want highly customized solutions, Sun is happy to oblige–that is 95 percent of the company’s business today. But, Sun is betting that transparent pricing based on some benchmark (such as processor load) is what enterprises will want, even if they don’t know it today. "Reliable repeatability is what drives scale and scale economies–not one-offs where every job is a custom software stack," Schwartz said.
Sun makes delivering a grid utility service and transparent pricing seem simple and straightforward. But many issues have to be resolved, such as establishing a measurable benchmark for comparing the market value of grid compute services. In Sun’s demo for a retail site, where you can log in and submit a job, there’s wasn’t an option to get an estimate of what the job would cost prior to unleashing the grid. Youngjohns told me that Sun is looking at various grid engines and working on an intelligent object interface, which would allow customers to drag and drop icons to initiate jobs.
Sun has a cost advantage because it owns most of the pieces required to power a grid, but a profitable grid is more than the sum of the hardware and software. Managing growth and amortizing tens of thousands of CPUs and equipment are fundamental to the economics as well. With more competition, margins come under pressure. It’s guaranteed that the pricing matrix will get more complicated, selling processor X for $1 per CPU per hour and processor Y for 70 cents per CPU per hour.
Certainly, there is a lot of flexibility in the pricing. I am sure AMD gives Sun a nice price break to help promote the growth of massive AMD server farms around the world.
Sun has a history of coming up with great technology, and competitors who use it to build billion-dollar businesses. Grid utility services aren’t a unique Sun invention, but the company is challenging the industry to establish commodity markets around computing power. The question is whether Sun can become the preferred infrastructure provider across a wide range of industries in which competition boils down to price per unit of compute power. It’s too early to tell, but this is the beginning of what is a trek to infrastructure-as-a-service.
