Tim asks

Besides cost savings, what are the benefits of virtualizing server, storage, and desktop systems?

Deane asks

I’ve heard cloud computing can save money. What is it, and where are these savings coming from?

Chris asks

What is the best way to compare the real costs of new IT equipment? For instance, I know some products may cost more but offer greater energy cost savings long term.

Herbert asks

I keep hearing that virtualization will save our business money. Is there any sort of calculator I can use to really see where and how we would be saving? How much does it cost to implement? and where do these savings actually come from??

“Physicalization” uses mobile tech to offer an alternative to virtualization

The server space has run into a problem. The continued operation of Moore’s Law is delivering greater amounts of computing power, storage, and bandwidth every year, and the demand for these capacities is escalating as well. However, our ability to efficiently make use of this capacity under one OS instance isn’t increasing as rapidly. This is aggravated by the trend toward multicore computing, which enables a level of parallelism that modern software still hasn’t fully accommodated.

The answer to this dilemma on the software side is the same as it is on the hardware side: increase parallelism by introducing a greater number of OS instances per processor, just like we’ve introduced more cores per processor. One technology for doing this is virtualization, which uses large, multicore, multisocket servers to run many OS instances. Virtualization has been on a massive upswing in adoption in the last few years, as this problem becomes more and more pressing. But management overhead is one of the by-products of virtualization, and great care is required to make virtualization work properly.  So some small datacenters may be looking for an alternative to or a complement to a full virtualization deployment.

Several server vendors are pushing an alternative solution that could be a better fit for some types of workloads; this solution often goes by the name of “Physicalization.” Physicalization involves building smaller servers out of low-power processors, tailoring them to the amount of computing resources which can be efficiently used by one OS instance, so that a one-to-one correspondence between hardware and instances can be maintained, thereby bypassing the management overhead and licensing costs of virtualization. As a side benefit, low-cost, low-power hardware, like the Intel Atom designed for netbooks, can be employed, rather than expensive server hardware.

A number of physicalization solutions have been announced in recent months, and while these solutions can’t offer greater density or PPW, they do represent a workable alternative to virtualization for some users. For instance, web hosters can offer their clients dedicated hardware, while using commodity hardware made cheaper by consumer economies of scale.  They also allow centers to harness the “free first gigabit” of ethernet over copper, and potentially avoid expensive 10GigE and Fibre installations. For some applications, physicalized servers may be the best option for using Moore’s Law to meet the ongoing challenges posed by increasing demands and limited resources in the datacenter.

It remains to be see, however, whether physicalization will actually catch on. The challenge that physicalization faces is that, for all its benefits, it simply doesn’t make as efficient use of Moore’s Law as virtualization does. Let me explain.

The problem with physicalization

Moore’s Law is most often cited as a statement about “computer power” or “the number of transistors on a chip” doubling within a specific timeframe (typically 12 or 18 months). This explanation is serviceable enough for the popular press, but it’s not entirely accurate. In its original 1965 formulation, Moore’s Law was about the cost-per-transistor dropping as a result of increased integration. In other words, as you pack more transistors into the same-sized area of die space, the cost per transistor drops dramatically given a fixed cost-per-chip.

Chipmakers take advantage of this declining transistor cost in two distinct ways: 1) they pack more transistors into the same-sized chip and charge customers roughly the same amount, but for more computer power, or 2) they make more chips that are smaller and cheaper, so that customers get the same amount of computer power but for less money.

The first option is typical of the server space, and the second is typical of the mobile space. But what physicalization tries to do is to take the second option into the server space. Unfortunately for the proponents of physicalization, if you’re going to put more hardware into the same machine, it’s always cheaper and more power-efficient to integrate that hardware onto a single processor die. This is because, as an iron-clad rule of computing, die-level integration is always cheaper than board-level integration.

So by taking multiple smaller, cheaper chips that are intended for separate mobile products, and using board-level and network-level integration to cram them into a single server product, physicalization vendors are using Moore’s Law in an odd, non-standard way that forfeits one of its major advantages, i.e., the advantage of die-level integration over board-level integration.