Marketing Memos

I am not the only industry observer to foretell the coming commoditization of server hardware. Many pundents have noticed that servers with PC origins have dwindling differentiation (Sun’s protests not withstanding).

While Linux and Windows sped the rate of WinTel/LinTel server commoditization, the ever shifting sands of tribal partnering slowed this process. This due in part to the persistent need for cluster or grid specific technologies to get commodity servers to scale for the more mammoth jobs found in IT and high performance computing (HPC).

This may be ending.

One of Silicon Strategies clients is Virtual Iron. Just out from under stealth mode, these folks are doing something heretofore thought impossible. They are gluing together multiple physical servers to look like one really huge x86 server, all without modifying Linux or requiring specialized applications (massive disclaimers apply, but are at best trivial).

Until now, x86 server hardware enjoyed moderate protection from full commoditization. Customers would buy from select vendors for a variety of reasons, not the least of which was the keiretsu of partners that provided an complete cluster or grid solution. Virtual Iron may well eliminate the need for such strategic alliances, and in turn drive x86 servers into true “commodity” status, where the primary differentiation is price.

Needless to say, HP, Dell and IBM should not be happy about this.

How can this be? The answer lies in the elimination of specialized software and hardware typically required to scale a cluster or a grid. Clusters require mass provisioning schemes, outboard load balancing tools, and other accessories to work. Grids require applications to be “grid aware” to make use of distributed resources. Customers bought into a fixed set of vendors to avoid the unpleasant side effects of “non-standard” implementations (a term which, if viewed in a cynical way, is an oxymoron). This allows server vendors to charge a moderate premium.

This may be coming to an end. The Virtual Iron product allows vanilla servers and blades to be combined like so much raw ore. Legos. Tinker toys. Bits and pieces that when bolted together create bigger and bigger virtual computers. No provisioning software needed (one instance of Linux per primary application would do for most shops). No grid-aware applications (any application that scales on an SMP system will scale on Virtual Iron). An if you run out of horsepower, bolt in another server and press the power button - it will be added hot to the virtual server.

And here is the kicker: Virtual Iron is not picky about matched components. All the blades in a rack do not need to be identical. Their can be different memory sizes, different CPU counts, and different clock speeds. Virtual Iron can balance loading appropriately.

If Virtual Iron grabs market share, the lack of need for hardware differentiation will drive commoditization further. Several vendors have produced “PC’s on a chip”, and used them as building blocks for specialized supercomputers. With Virtual Iron, we are a stones throw away from using the same reduced hardware as the foundation for “personal supercomputers”.

HP, Dell and IBM should be worried. By diminishing the value of the standing cluster/grid keiretsu, Virtual Iron jeopardizes the last remaining differentiation for x86 servers. Vendors will have to shift more towards standardization, more towards miniaturization, and more towards price (and margin) cutting. Good for the consumer, bad for the vendors.

Server makers should also be fearful of the lost revenue from the end of  “over scaling”. Every systems administrator on the planet overscales every server in their data center to meet peak loads. But it is a rare day when all the servers reach peak loading at the same time. Virtual Iron allows for the creation of more than one “virtual computer”, and allows hardware to be shifted between them - hot, without bringing Linux or the application down. A savvy shop could avoid overscaling all their servers, and thus spend less on hardware, driving down HP’s, IBM’s, and Dell’s revenues.

And it gets worse. Since Virtual Iron is indifferent to the capacity of the servers being added, and allows servers with differing configurations to be merged together at will, upgrade demand is reduced. IT shops will have less need to decommission or repurpose old servers. Instead, they just add them to their virtual computers and use their hardware until it dies. Less volume for HP, IBM and Dell.

But, that is not the end of the story. Assuming that Linux can scale beyond current limitations, the Virtual Iron technology may well cause erosion of high end systems (are you listening, Sun? ). If in the near future one could bolt together a string of Opteron systems that (as far as Linux is concerned) looks like a really big SMP box (say, rivaling Sun’s top Sun Fire servers ), then margins for top-shelf servers erode as well. If one can install Linux on a 128 way virtual Opteron server, and you can compete HP, IBM and white-box vendors against one another, Sun adds no value. Nor do IBM zSeries systems.

So, where will this all end? True commoditization. As long as the servers or blades being added to a Virtual Iron environment are x86 systems (64-bit is coming), then they can be used. Since I/O and storage is virtualized, there will be little need for internal differentiation. The final evolution may well be the reduction of a “server” or “blade” to a chip with an InfiniBand or multi-gigabit Ethernet port, and nothing else. With an utter lack of differentiation, “servers” could be manufactured en masse, sold in six-packs, and have prices fall through the floor.

Let’s hope.

Bookmark to:

Perhaps I just don’t get it, but I cannot find the value behind Sun’s Rent a Grid.

With more fanfare than today’s Superbowl, Sun announced the availability of their grid for hire. Sun’s strategy is to make computing horsepower like an electrical utility - you buy as much juice as you need, and only when you need it.

In their recap of the announcement - and unrelenting public bashing of the competition - Sun admits that their initial market is limited, but allegedly growing. In the hear-and-now, the market consists of high power users that have non-perpetual needs - ” . . . in our crosshairs, from risk analysis to movie rendering, data warehousing to reservoir simulation.” But Sun thinks the market will grow because ” . . . we also know that’s where the network’s headed - that more traditional apps are spilling into the grid, and the market’s growing.”

What weakens this near religious belief is that the rest of the industry is heading in the opposite direction. Instead of creating competing electrical companies, IBM, Dell, HP and others are delivering affordable power plants for the back yard. And with the advent of some incredible virtualization technology, customers will have their own mini nuke facilities.

So where does Sun’s advantage lay? The only ready buyers will be those who can model grid-aware applications, then deploy them for short periods of time. Digital animators like Pixar might be suitable, as well as some private scientific research facilities. But few others will buy since their computing needs are more frequent and recurring, and because the bottom continues to fall out of the hardware market. A lovely intersection of need, speed and greed.

What is the difference is between a utility and a commodity? A utility converts commodities into more useful services (an electricity provider converts radiation into alternating current). Since the cost of building an electrical generation plant, mining and converting raw materials, and monitoring the systems 24X7 is huge, very few people build their own power plants.

But what if those steps were already covered? What if the uranium was already enriched and available at the corner store. What if you could buy your personal nuke plant at Home Depot and it included self-monitoring capabilities? And what if Junior - that bright young kid of yours - knew how to assemble and operate the reactor? Then your set-up and operational costs would be low enough to make energy self-sufficiency a Good Thing ™.

That is precisely the case in IT today. IT has the trained staff. Vendors provide cheap raw materials (commodity servers, blades and Linux). And technology is packaged such that it requires less and less tinkering to get it all working well. Given the deepening commodity nature of the industry, buy vs. rent decisions quickly devolve to price comparisons.

And Sun’s Rent A Grid program ain’t cheap.

Sun’s charging $1/CPU Hour. Pretty good if you use their grids a few times a year and want to avoid the capital expense and setup of your own data center. But even in a worst case scenario, it is much more expensive.

I scouted around the net and priced blade servers. Adjusting for the number of processors, and assuming that analyst are right in that for every dollar ($1) of hardware you spend seven dollars ($7) in manpower and software, the math looks good for buying and not renting. My spreadsheet shows that the equivalent hardware at the inflated 7X rate would buy 1.6 years on Sun’s grid.

Given that this is below the traditional three-year breakeven point many IT buyers use for decision making, then Sun’s Rent A Grid is over priced by a factor of two. If we take the lesser cost - just hardware - then for the price of your own hardware you would get a mere 2.8 months on Sun’s grid.

Once buyers do the math, Sun’s plans may just become gridlocked.

Bookmark to: