March Issue of the Taos Newsletter: Utility Computing

Convergence: The Next Ten Years
Harry Quackenboss, President and CEO, 3UP Systems

Introduction

The emerging era of utility computing will bring with it major changes in the power of computers and the way people operate and use them. IT departments, service providers, and systems vendors, and component suppliers will have to adapt in order to meet the needs of their customers and keep abreast of the competition.

It Takes a Long Time to Become an Overnight Success

"The concept of an information-processing utility poses many questions. Will the role of information utilities be sufficiently extensive and cohesive to create a whole new industry?" 1

Utility Computing is a lot of things, but one thing it isn't is new. In the mid-1960's, some of the finest minds in computing got together to build the foundation of the computing utility. 2

While the vision of anytime anywhere access to computing has been longstanding, affordability of the computing power and the pervasive high-bandwidth communications needed to fulfill the vision is only now becoming reality. What does this mean for all of us in this industry?

It's likely to bring about unprecedented change. It might even bring about a major re-alignment of the major players as dramatic as what happened between 1985 and 2000, when almost all of the incumbent leaders were displaced. This may sound like fantasy in the era of today's behemoths. But so did predictions in the 1980's that relational database management, UNIX and merchant microprocessor-based computers would dominate the computer industry.

How times change. Then again, maybe not.

The past 20+ years have been about the explosion from centralized to functionally distributed systems with separate subsystems for servers, networking and storage. This brought with it the destruction of almost all of the mainframe suppliers, the decimation of the minicomputer industry, and the rise of merchant microprocessor servers and desktops, with operating systems, data base managers, networking and storage subsystems from different vendors.

In the current ecosystem, the servers, networking and storage products are, for the most part, designed and produced by separate companies. Presuming for the moment that things happen faster now, let's look at the next decade.

It might turn out to be about implosion, or more gently stated, convergence. Let's examine some of the influence factors.

Hardware Is the Driver

Like compound interest, Moore's Law adds up. Over 10 years, it amounts to a 100X increase in the number of transistors on a chip (Figure 1). Like a black hole, this ever-increasing chip capacity sucks in surrounding stuff. Boards collapse into chips, chassis get sucked into boards, racks get sucked into chassis, and so on. This is true at chip level, at board level, at box level, at rack level, at data center level, and, at least for the high technology hardware and software vendors, at company level. The impact of Moore's Law, over time, drives whole new architectures.

X86 Servers Are Changing the Landscape

The server arena is about to see a big increase in the use of X86 processors from Intel and AMD. For most of the last twenty years, the processor performance of X86 instruction set processors was at a significant disadvantage to RISC processor designs from IBM, HP, and Sun, etc. While the leading RISC processors may still be faster than their CISC counterparts, the difference isn't the 2+X advantage it used to be. It's not just integer performance either. Thanks to the appetites of home gamers, these CPUs have built-in vector processors with impressive floating-point performance. 64 bit extensions from both Intel and AMD, will increase the acceptance of X86 for big (a relative term) systems.

Dual-CPU Servers Will Dominate

The performance of X86 CPUs and clustering are enablers for another phenomenon, driven by economics: servers with 2 CPUs will dominate the multiprocessor server category. This starts with silicon. Some of this is in the CPU itself, but a lot of it is inside the chipsets that connect the CPU to peripherals, memory and other CPUs.

When it comes to chips, the highest volume parts get the most R&D, so silicon for 1 and 2 CPU servers will be on a better price/performance curve. The evolution of these product lines will also be faster, over time widening the gap. This ripples through the chain from boards to boxes. At a systems level, engineering, debug and regression testing of dual-CPU servers can be done faster than systems with more CPUs and complexity.

The New Age of Big Systems

Obviously with the push behind utility computing and its relatives (on demand, grid etc.), the world will need lots of big systems. Increasingly, 2 to 4 processor building blocks will be lashed together to mimic big systems. Innovation all over the place is directed at making this possible. It's happening with interconnect silicon, physical packaging, operating systems, networking, storage and systems management.

Linux clustering has been used for several years in science and engineering applications, but IBM and Oracle are now promoting it as an alternative to large SMP (symmetric multiprocessing) systems.

But won't this be inefficient compared to shared-memory multiprocessor systems? From the outside view, the goal here is to optimize the overall economics. It's all about balancing tradeoffs. The lower hardware cost compared to big SMP systems should, for most applications, more than offset the "waste" and "inefficiency" of the multiple copies of some of the data, and the communications resources used in moving stuff around.

But even from the inside out, in some situations, loosely-coupled systems might actually be more efficient than one big CPU. If the tasks are partitioned into small enough pieces that fit into the on-chip cache, there can be big performance gains. By partitioning tasks and sending the pieces off to several CPUs, each sub-task might run to completion without being swapped out, eliminating the overhead of the processes swapping in and out. While most applications developers won't go to the trouble to figure this out, a DBMS vendor might.

The Impact of Blade Servers

If Dual-CPU servers become the dominant server design, then modular chassis with single line card servers (blades) and integrated switching make a lot of sense. Like most things, this is will go through several evolutionary phases. The first of the modern blade servers were introduced in 2001. These systems use low-power CPUs, and are most often used in web server farms. These systems are typically 3U (5.25" rack units high) designs, and have limited configuration flexibility. The more successful products have integrated Ethernet switches, either 100Base-TX, or Gigabit Ethernet. These systems reduce cabling, and pack more servers per rack unit than other rackable designs. But, the low-performance CPUs, and limited peripheral options keep them from being broadly used.

In 2003, several vendors introduced larger systems, typically 6U or 7U (10_" to 12_" high) using Dual-XEON processor servers. They include built-in GE switches, KVM (keyboard/video/mouse) switches, and feature built-in chassis management to control and monitor the servers. This new class of servers is still in its infancy but is starting to be used in data center applications as an alternative to general-purpose stand-alone rack mount servers. These blade servers are running Microsoft server operating systems and different versions of Unix and Linux.

IBM and HP are the early leaders. It is worth noting that both vendors farmed out the integrated GE switch components to a Taiwan supplier. For every server blade, there are 2 to 4 LAN switch ports bundled within the blade server. This means that if blade server shipments reach 1+ Million blades in 2005, there will be 2 to 4 Million GE switch ports with them. Will this cause a significant realignment of the switch industry? The major switch vendors are likely working on what to do about this, but so far, they are not telling.

The blade server phenomenon is a good example of the effect of Moore's Law. In 1993, leading edge data centers used 4 Intel 486 servers per 19-inch rack, Today, high-end blade servers allow 80+ Dual XEON servers in the same space, including the aggregation switches.

It doesn't stop at racks. Consider the data centers that were painstakingly architected a few years ago with just the right power and air conditioning for the servers, network gear, and storage farms, for what was then perceived to be, the future. Many will have to be remodeled to take optimum advantage of blade servers.

The Storage Angle

Along with servers and networking, online storage is the other corner of the data center triangle that fits into this converging world. Today, integrated storage options for blade servers are typically limited to two disk drives per server. During this calendar year, we will see the introduction of new storage options for blade servers that will package enterprise-grade storage subsystems in blade server chassis, with the flexibility of more drives per server.

In big data centers, storage farms may stay physically separate, but in smaller enterprises and branch offices, integrating servers, networking and storage into the same remotely-managed chassis will become an attractive option.

2004 will also see another storage option for blade servers. No Disks. While diskless Linux & Unix servers are relatively common, the only option for diskless Microsoft servers has been Fiberchannel. New storage subsystems are starting to appear using iSCSI. In a nutshell, iSCSI is storage connected via Ethernet, that looks to the server like a local drive, rather than a network drive. At least one vendor has introduced a capability to boot a Microsoft server operating system from an iSCSI storage volume, eliminating the need for local disks.

Software Evolution at Glacier Speed

Well, not really. But, compared to hardware, the rate of software evolution seems so. Consider how long "modern" software technologies have been around (Figure 2). Hardware evolves the software faster than software evolves. Software isn't the reason for mobile phones running Linux. It's cheap, fast silicon.

New Software Requirements

Utility Computing, just like the electric and the telecom utilities, will benefit users with lower costs, if utilization can be improved. Demand will have peaks and valleys. Some of it is seasonal, such as around year-end reporting. Some of it is caused by time zone differences and some of it is unpredictable. Plus, an always-on resource means dealing with both scheduled and unscheduled system maintenance. Better tools to dynamically reconfigure the utility should improve utilization.

There is another lesson from the utilities. Just like the modern telephone service providers have delegated responsibility for call setup/teardown, conference calling and voice mail management, to the end user, meeting the promise of Utility Computing will require new tools, procedures and billing mechanisms to delegate authority to user organizations to reconfigure their resources.

Opportunities for New Vendors

It's not just the servers that need to be reconfigured. The configuration of the network, VLANS and storage needs to be coordinated in real-time with the servers, along with access rights to all the various elements. Even when the various subsystems are in separate boxes from different vendors, common management is needed. This is no easy task. In this area of rapid change, standards for the various interfaces among all these pieces are years away from being widely available.

Since real-world usable standards will take a long time to translate to products, the vendors who can provide converged solutions including servers, networking and storage management may be the winners over the next decade. The incumbent vendors have compartmentalized. The server vendors have moved from being developers of their own communications product lines, to resellers. How they evolve in this new converged world will be worth watching. If the past is an indication, some of the winners will be new vendors that displace incumbents.

Implications for IT Organizations

Ultimately, in this distributed industry, the burden falls on the enterprise IT departments to make sure that the particular combination of hardware and software, under their particular user load, works. All of this means an increase in complexity of the architecture and planning of enterprise systems. IT organizations are likely to rely increasingly on outside expertise for planning and deployment.

Among the factors that are important to IT organizations in this movement to the utopian future of Utility Computing, one stands out. The lines in IT organizations that separate server management, networking and storage are going to blur in this new era of convergence. It's not too early to do some cross-pollination, maybe mix up the organizations a bit, in order to make sure each area understands the other.

Increasingly, the end users will demand always-on operation. Many IT organizations, especially below the Fortune 100, will find it more cost-effective to outsource ongoing support to organizations staffed to provide the 365 by 24 hours per day services needed.

Harry Quackenboss is president & CEO of 3UP Systems. Contact Harry at quack@3upsystems.com

1 "The Computers of Tomorrow" by Martin Greenberger, The Atlantic Monthly, May 1964 http://www.theatlantic.com/unbound/flashbks/computer/greenbf.htm

2 See http://www.multicians.org/fjcc1.htm