THE DOLLARS AND SENSE
OF SERVER CONSOLIDATION		    
 

The next round of server consolidation may provide the biggest price/performance disruption in the IT market since RISC.

    
  by  Jack Fegreus      
     
 

In the late ’80s and early ’90s, business consultants were extolling the virtues of the distributed organizational model. Corporate hierarchy was synonymous with a lack of empowerment, inflexibility, and inefficient bureaucracy. This in turn had far-reaching effects on IT, which exists to support the business process model. As companies reengineered core business processes on decentralized business models, IT adapted around a distributed, decentralized computing model, too.

The distributed, decentralized business model was initially attractive because of relatively low costs associated with the incremental acquisition of capital and labor. On the other hand, excesses in the new business model emerged in increased long-term people costs, support costs, and administrative costs. For IT in particular, supporting the distributed business model meant a proliferation of servers throughout the enterprise.

The initial value propositions of a distributed computing environment centered on low incremental acquisition costs and a supposed advantage in the elimination of application development bottlenecks inherent with centralized IT. What was not factored into the equation was that too many distributed servers means slow response times due to LAN bottlenecks, as well as varying consistency in security polices and—even worse—the data itself. All of that adds up to the introduction of serious problems in areas such as customer and user service. That’s the rub in 2002.

The new mantra of business-organization gurus is now ‘The Customer Experience.’ Good bye ‘empowerment and creative problem solving’ and hello ‘service, service, service.’ Added operational costs are the least of your worries today, if you have a business model that leads to service problems. At a time when building new internal cross-application services and introducing customer-friendly web services are seen as priorities, data inconsistency problems that prevent the development of these services dwarf any recollection of an application backlog.

The result has been a revived focus on Total Cost of Ownership (TCO) and improved efficiency for IT. And what consistently turns up in every IT model from Gartner to IBM is the fundamental fact that operating costs dwarf acquisition costs for servers. In particular, a June 1999 report by the Robert Frances Group pegs the acquisition cost of a server as representing only 20% of the TCO. These increased operating costs come from both the well-known sources of network, security, and operational management to the less well-known source of poor hardware resource utilization (see sidebar).

 
         
 
IT'S ALL ABOUT U

by Chet Heath, VP and CTO, OmniCluster Technologies
Web hosting services are often located in metropolitan areas where the cost of floor space is high. To conserve space, these servers are typically installed in floor to ceiling racks, arranged row upon row, perhaps resembling crops arranged neatly on a farm. The size of the server, how much space it consumes in a rack, how much power it consumes, how much heat must be removed from the room by air conditioning, and even how much electrical mains infrastructure is required become major cost items.

Gartner Group has done extensive research on the total cost of computing for desktop situations. In enterprise situations, desktops outnumber servers 28 to 1, so it is logical to focus on the desktop there. In web hosting situations, the ratio is reversed with likely 100 times as many servers as clients; so the same assumptions cannot apply. Indeed, one of the largest factors in desktop TCO modeling, user support for application software installed in the computer, does not really apply in the hosting world. So there needs to be new cost assumptions.

The business model of a web-hosting establishment, above, can be simplified to revenue coming from one source: Servers providing web pages. Customers pay to rent the use of a machine, or part of a machine and that revenue is used to pay overhead and recurring expenses. The web hosting business is an intelligent machine based model, very much analogous to the human based direct sales organization.

Costs are similarly attributed to the machines and the positions that they occupy as machines outnumber humans by a large factor. From interviews with actual web hosting / collocation businesses, this is what the big factors are fixed overhead expenses and scaleable recurring costs:

Fixed Overhead Expenses

A. Real cost of the space in the rack. This isn’t simply the cost of a rack divided by the number of “U” space units (distance in repeating mounting hole pattern) consumed on the mounting face of the rack by the server. It has to include all of the real support costs for the rack and installed computers as well. Further, it has to include the fixed costs to set up the configuration, like the cost of electric mains, delivery and installation into the rack. Attributing the costs per “U” is therefore very much like attributing costs per desk as overhead expense in a human based enterprise.

What these costs are varies a great deal with the fanciness of the configuration. The author knows of one “Top Tier” hosting enterprise that buys everything new, has elegant racks with glass doors that showcase the computers and is located at a premier address. While another “discount” web-hoster buys open racks on surplus, builds everything themselves and finds an industrial district to be satisfactory as a home. The discount model does not scale well with volume, as there are only so many used racks, power generators and low rent homes available. There is one thing in common between both businesses: rack space is a major factor. There are typically 40 to 42 “U” space units in a rack. Because of fixed costs that must be attributed to the servers as a source of revenue, it is some of the most expensive real estate on the planet!

B. Cost of the Server complex. Some of these are non-intuitive extra expenses. It costs something to deliver, set-up and cable the server. While some configurations enjoy the quality of Tier-1 solutions with sales commission, warranty and support factored into the price, others may be assembled from surplus or used computer parts and supported locally by the user. Tier-1 manufacturers are quick to point out the “Pay me now or pay me later” aspects of building your own machines. The Fortune 500 enterprise world has come to agree with that. Since there is no way to create a uniform yardstick with home-built “white box” systems, Tier-1 machines are assumed in the TCO factors here.

C. Software licensing OS and applications. It is assumed that the license fees are paid, although Linux and BSD are open source licensing with primary support provided internally with external support available. Windows is moving to a “leasing model” perhaps to make comparison to the Linux model more direct. This is a platform dependent (almost cultural) choice, independent of infrastructure and cannot be addressed in this article.

Scaleable Recurring Costs

A. Web access fees

B. Electric Power for lighting, air conditioning, operation of the servers, support of the office staff. Also Maintenance of Bldg, generators, interior fixtures, servers

C. Support Staff, maintenance, upgrade, replacement personnel

Reader’s Cost Verification

Empirically, the non-electronic web hosting cost can be verified by contacting web hosting and co-location enterprises and requesting a rental cost per “U” quote in existing racks. A standard assumption of a return on investment in 40 months, on 10-year depreciation cycle, would indicate that the rental should be about 3 times cost. This type of markup is required to cover technology obsolescence and the scaleable recurring costs: Server equipment, which may be replaced every 5 years or so, maintenance on the backup generators ($4K/yr typical), web access fees $3/megabyte (not untypical), plus salary and benefit costs for human resources.

Cost Matrix per Rack
Fixed costs exclusive of electronics

Tier-1*
More than 1000 racks

Discount
(161 racks)

Cost of Rack
 Includes delivery

 $660 wholesale $215 used
Space allocated to servers/racks** 140 sq ft 140 sq ft
Electric infrastructure per rack 4800 watts 4800 watts
Electric mains wiring per rack $2000 $1005
Raised flooring $300 $204
Real Estate Value/rack
assumes a 10yr depreciation of current values		 $250/sq ft = $35,000 $141/sq ft = $19,700
Air conditioning $1000 $607
Backup generators and transfer switch wiring $276 288
Internal building construction $300 $178
Electrical engineering design $150 $28
Other inspections and fees $200 $124
TOTAL nonelectronic infrastructure $39,886/rack $22,349/rack
 TOTAL per U (42U per rack) $950/U $532/U
 

One of the time-proven strategies for those attempting to minimize TCO for IT is simply to reduce complexity whenever possible. Here a common tactic is just as simple as the strategic concept sounds. Sites embark on a plan of physical consolidation under which the absolute number of servers is reduced but within the same hardware and software architectures.

Tactics can be as simple as reducing the number of servers, storage devices, and the number of associated vendors. The one thing in common for sites that implement a successful physical consolidation strategy is that their key drivers are always the applications that are being run. This is why IBM refers to server consolidation not as a solution but as an enabling technology.

One interesting effect of physical consolidation is that often there is a serendipitous improvement in application performance. A study by IDC of physical consolidation of IBM iSeries servers (AS 400s) reports that consolidating five AS/400s into one AS/400 yielded a reduction of almost 15% in support costs and simultaneously increased application performance by 90%.

Another form of consolidation that arises by making the applications and not infrastructure the critical drivers of the process is dubbed data consolidation. Here the tactics are to combine data from multiple sources into a single repository. The most immediate results from any successful data consolidation plan are reduced storage management costs, improved backup and recovery capabilities, and •lower administrative costs.

More importantly, data consolidation can greatly improve both data access speeds and data integrity. These latter two benefits are essential for enabling the internal cross-application services currently high on the IT-needs list of any corporation. What’s more, they also are critical for future customer-oriented Web services.

Physical consolidation of servers, however, is only the tip of the iceberg. The new wave of consolidation has been dubbed by some as operational consolidation. Under this scheme, physical machines exchanged for logical machines on a host server. More often than not, operational consolidation involves the introduction of a new computer architecture. It is here that the advantages of Open Source applications and operating systems come to the forefront.

IBM recently announced plans to deliver the industry's first dedicated Linux-only mainframe. The IBM eServer zSeries for Linux opens a new chapter in IBM’s Linux strategy by addressing an entirely new class of customers with no traditional mainframe operating system skills using IBM’s z/VM virtual machine technology. With just one IBM eServer zSeries system, customers can consolidate from 20 to hundreds of stand-alone servers. In Finland, the telecom Sonera Entrum is partitioning an IBM eServer zSeries system into 500 virtual servers each running its own copy of SuSE Linux.

IBM also announced plans for a Linux server for small and medium-sized businesses, IT's fastest growing market segment. Leveraging IBM's advanced logical partitioning technology (LPAR), the iSeries architecture provides scalability, availability, and high-speed connectivity for OS/400, Linux, Java, Domino, WebSphere, Unix, and Windows 2000 Server applications. Customers can consolidate up to 31 Linux partitions onto a single physical server. The eServer iSeries supports Linux distributions from SuSE and Turbolinux and includes an installation wizard for rapid deployment.

Rapid creation and deployment of logical partitions is the crucial element introduced with the operational consolidation equation. Server farms with endless 1U racks of servers are no longer just the province of ISPs and ASPs. Many large corporations have found this processing model to work for them as well. They key advantage that logical machines bring is an end to having to stage a new server. Simply copy and clone an existing logical server and in less than five minutes, a new virtual machine is up and running.

That is of course assuming you are running Linux. Do that with Windows and expect a knock on the door from the licensing police. As far as Microsoft is concerned, virtual machines require very real licenses. And making this point all the more important is the burgeoning movement to bring operational consolidation to the PC market via server blades. Essentially single-board CPUs that plug into a PCI bus, server blades hold the key to the most dramatic price/performance drop for PC servers that IT has ever seen.