A marginally different version of this article appeared on as of October 22/01. Please do not re-publish without their permission but feel free to make copies for friends or personal use.

A strategic comparison of Windows vs. Unix

Most of the Windows versus Unix debate has been cast in terms of which is technically better or which is cheaper but the real question is, "Under what circumstances is it smarter to pick one technology rather than the other?" (4,500 words)

Paul Murphy is a 20 year veteran of the IT consulting industry and the author of The Unix Guide to Defenestration

In this paper we examine the Unix versus Windows decision by asking which is the smarter choice in each of two situations:

  1. in the first one a college administrator has to choose between a Unix and a Windows architecture for a 500 student system; and,

  2. in the other we look at that same decision from the perspective of the board at a 5,000 user manufacturing company.

Under each scenario we try to consider both the direct and indirect consequences of the decision. We look, for example, at the impact the college's decision has on parental support costs for students attending the college and not just at costs incurred by the college. Similarly, for the manufacturing company we look beyond cost and consider the impact the board's decision has on things like user productivity and the CIO's role in the company.

Case 1: Student system

We picked the college situation because its requirements are relatively simple with few management issues to consider. The administrator wants to minimize costs while giving the students access to standard tools, including word processors, Web browsers, and e-mail. The school also needs to allow staff to communicate with each other and students using Web servers, shared files, and e-mail while avoiding security or other systems-related crises that would detract from the business of the college --teaching.

Cost data used in this article was collected from the Microsoft, Dell, and Sun websites as well as by email from NCD on or about September 19th, 2001.

In an all-Windows situation, those requirements spell out a need for a rackmount of four small servers and 500 desktop PCs. All would run a version of Windows 2000 with Back Office on the server and Office 2000 on the desktops. The installed cost to the school, using Dell Poweredge servers and Optiplex PCs, come to about $1 million.

The at-home cost imposed by the college's decision is similar in structure except that the parents don't directly pay for servers while security and other support issues are absorbed in uncosted student time. If students share about 20 percent of the machines, this amounts to a requirement for another 400 PCs, in this case Dell Dimensions 2100s with 15-inch screens, Microsoft XP with Office 2000, and the 256 megabytes of memory needed to run it. Ignoring communication gear and related issues, these cost the parents about $574,400 at start-up.

As a result the college's Windows decision has a total capital cost of about $1.6 million.
Hardware Software Unit Cost Total Cost
(504 users)

At work:
17-inch Dell GX150; 128 MB of RAM, 20 GB hard disk, 900-MHz Celeron

Windows 2000
Office 2000
At home:
15-inch Dell 2100 with 256 MB of RAM & network card
XP/Office $1,436 400 users because of some sharing
Rack of four Dell Poweredge 2550; 2 x 1-GHz PIII, 2 GB of RAM, 36 GB disk, dual controllers
1 TB PowerVault shared storage
Windows 2000 Advanced Server
Back Office 2000
499 CALs
School Total     $1,059,980
Parent Total     $574,400
Start-up total     $1,634,380

What's a smart display? It is a desktop graphics device intended solely to handle the user interface components of an application running on a server. Most provide and extend the capabilities of such predecessors as the "dumb" terminal, X-terminals, and the Microsoft "thin client" while adding Java/OS and local browser execution capabilities.

In the Unix alternative, the college meets student needs with a Sun 4800 server, an administration workstation, and 500 smart displays running bundled software, including OpenOffice, Netscape, and the usual suite of Unix e-communication services for a total cost of about $630,000.

Parents will still need to provide a home computing device of some kind. Here we assume parents and students would choose a Linux or BSD-based PC using the same set of free software the college chooses. This reduces the student's need to re-learn applications and operating systems while protecting the student's ability to read and write Microsoft files for interchange with people at other schools or businesses.

The immediate software savings, relative to the Windows alternative, would allow them to buy both a better quality machine and a larger screen for less money. Choosing, for example, a 17-inch Dell Optiplex at $1,219 plus $50 for a Caldera CD costs about 12 percent less than Windows on the 15" Dimension 2100s for a total cost of about $507,600.
Hardware Software Unit Cost Total Cost
At work:
500 x 17" SunRay smart displays and one Sun 4800 with 12 GB of RAM, 1 2 x 750 CPUs, 2 TB SCSI Disk
Solaris with all needed application software $621,470 $621,470
Administration workstation: SPARCstation 10; 512 MB of RAM, 2 x 20 GB disk, CD-ROM & floppy drives, 21-inch monitor Solaris $5,345 $5,345
At Home:
Dell Optiplex 17-inch with Caldera Linux
OpenOffice and related open source tools $1,269 $507,600
Cost to College     $626,815
Cost to Parents     $507,600
Start-up total     $1,134,415

When we look at five year costs for these two configurations, three things stand out:

  1. 24 x 7 reliability and freedom from the effects of student attacks on system integrity are givens in the Unix environment and virtually impossible to achieve in the Windows based one;

    For schools, that's a key selling point - and the reason Sun put over 100,000 of these units into schools last year - there are no desktop vulnerabilities and even disassembling a smart display doesn't get the student anything.

    Some faculty will demand that their favorite Windows-only software be made available to students. If necessary this can be done via embedded processors (up to 8 in a Sun server) running NT/Win2K, by attaching a Windows server to the network, or, for small scale solutions, through use of a Windows emulator.
  2. the Unix administration job is really part time although, in practice, it would be filled as a full time position and the person hired will find additional ways to contribute to the college. The Windows based solution, in contrast, will be under supported with four full time staff and lead to a serious loss of productivity among other professionals as they become part time PC support people; and,

  3. In the Microsoft world incessant product churn is a given with each new generation embedding new ideas, and new technical directions, that make it incompatible with the previous one.

    For example, someone who bought Office 95 for Windows 95 on a Pentium 133 with 32MB on release of the product in 1996 would, five years later, have had to upgrade his PC at least once and his software at least twice before being forced to replace all of it - just to ensure that he could continue to work with files sent him by others. This churn rate shows no sign of slowing. Today, that same Windows buyer would be facing yet another forced march, his third in six years, to new hardware and software to cope with the XP/net generation.

    That doesn't happen in the Unix world. Unix is built on solid theoretical foundations; technical progress is faster, and goes further because the next generation extends, but does not replace, the current generation. That's why the Kernighan and Pike book: The Unix Programming Environment (Prentice Hall, 1984), is valuable to today's Linux user but last year's Windows/ME book is useless to someone buying Windows/XP.

    As a result parents buying a 1.1GHz Dell PC with XP today can expect to be buying at least two further generations of software and a new machine before the student graduates, while those whose kids use Linux will see them taking their original machines along when they leave home.

Product churn is a major contributor to the long-term direct cost of ownership for Windows users. That, coupled with staffing demands, is why its 33 percent cost disadvantage relative to Unix at purchase time grows to more than 66 percent over the first five years.

Cost Source Windows Cost Estimate Unix Cost Estimate Percent Savings
with Unix
Initial Capital Cost   $1,634,380   $1,134,415 31%
Support Staff 4 x 45,000 x 5 years $900,000 1 x 65,000 x 5 years $325,000 64%
24 month software refresh 2 x 479 x 504
2 x 400 x 349
36 month hardware refresh 904 x $1,219 $1,101,976     100%
Parent Total   $1,341,209   $507,600 62%
School Total   $3,057,179   $951,815 69%
Total Estimate   $4,398,388   $1,459,415 67%

Case Two: 5,000-user manufacturing operation

The 5,000-user manufacturing company example is more complex, not because of hardware or software issues, but because this comparison gets into the more important, but largely unquantifiable, areas of the overall value of systems to a company and the managerial impact of a change in systems technology.

Since we don't have a good source for estimates of the cost of the PeopleSoft suite -selected for its client independence- we assumed it to be the same for both architectures. This probably isn't right, DB2 for Solaris is, for example, much cheaper than SQL-Server for a rackmount of Compaq Proliants and, of course, the whole client licensing thing doesn't exist for the Unix alternative.

That said, the Windows capital cost comes to about $11 million or $2,157 per seat before application software, wiring, installation, communications, or additional PC and server software licensing.

Hardware Software Estimated Unit Cost Estimated Total Cost
17" Dell GX150; 128 MB of RAM, 20GB, 900-MHz Celeron, Windows 2000
Office 2000
(5,200 users)
Set of eight in four racks. Compaq Proliant 8000, 8 GB of RAM, 8 x 700-MHz Xeon, 4 x 36 GB internal on Dataguard controller; 2 external smart controllers, R6000 UPS
Compaq 42U rack
Compaq rack monitor
Windows 2,000 Advanced Server
Microsoft Operations Manager with Application pack for 64 CPUs
8 x $86,923
4 x $5,439
4 x $1,625
4 x $2,500
64 x $1,798
Storage Works 4354R - 12 drives, mirrored 72 GB of RAM [net = 432Gb] + 2 DLT tape drives (110/220GB)   8 x $39,995 $319,960
Communications and related Estimated as Back Office with 5,200 CALS 4 x $3,999
5,195 x $232
Start-up total     $11,219,808


We envisage two data centers for the Unix solution. Each has separate administration and networking, with two Sun 6800 servers and 4 terabytes of shared storage for each pair. In normal operations each data center will run roughly half the load, spreading their share across both machines to provide 60 percent more cycles than the Windows alternative while running nearly everything at memory speeds.

This architecture is doubly redundant. One machine in each data center or one data center can be shut down without stopping, or even seriously affecting, corporate processing.

We use 21-inch NCD NC900 smart displays for the 2,000 or so office workers here instead of the 17-inch SunRays because these offer faster graphics, more screen space, and higher resolution.

Hardware Software Estimated Unit Cost Estimated Total Cost
NCD NC900 (21-inch monitor) NCDware 2,000 x $1,825 $3,650,000
SunRay (17-inch monitor) Solaris 3,000 x $600 $3,568,000
Two sets of 2. Sun 6800, 24 CPUs, 48 GB of RAM,
4 x 36, four controllers
Solaris, includes all needed tools and applications 4 x $892,000 $3,568,000
Two x T3 Storage Array with 5.2 TB of disk and 1 GB cache per controller   2 x $473,500 $947,000
Start-up total     $9,965,000

This structure produces an initial capital cost estimate that's about 12 percent less than that of the Windows solution. This system, however, offers larger screens, far more processing power, and is doubly redundant while the Windows system isn't redundant at all.

In operation, a Windows-based client-server system of this complexity is going to be staffed at a support ratio of about 30:1 and so will need about 165 full-time support people plus a base staff of about 35 for single shift operation.

The Unix system, in contrast, needs no more than 20 people on each data center team reporting to a CIO with a staff of perhaps five for a total complement of 45 or fewer to achieve full 24 x 7 operation.

Ignoring maintenance, space, power, and telecommunication as minor issues, long term costs stack up as:

Cost Source Windows Cost Estimate Unix Cost Estimate Percent Savings
with Unix
Initial Capital Cost   11,219,808   $9,965,000 12%
Support Staff 165 x 45,000
35 x 60,000
45 x $65,000 $14,625,000 63%
24 month software refresh 2 x $3,585,000 $7,170,000     100%
36-month hardware refresh   7,263,972     100%
Total Estimate   $64,878,780   $24,590,000 62%

for a total that's a bit more (38%) than a third of the comparable Windows cost.

The impact on home computer users is, of course, about the same per user as that for the college and works out to about $1,735 per household for the period. Assuming that all 2,000 office workers have home systems and switch to Linux (or BSD), the corporate decision to use Unix with smart displays would save these employees a cumulative $3.4 million, or 53 percent of the Windows alternative.

The productivity paradox & technology choices

The quality of board decisions should show up as changes in the return the company offers its shareholders. At the company level such clarity is relatively rare. Although failures like Hershey's ill-fated SAP implementation get a lot of publicity, most of the effects of these types of choices by individual boards are hidden from public view by obfuscating factors like market change, personnel change, or technology change in an industry.

One place the effects do show up is in aggregate economic statistics because these average out both intra and inter-industry effects over time and across sectors. In 1987 Robert Solow, a Nobel winning economist at Stanford, pointed out that "You can see the computer revolution everywhere except in the productivity statistics" and thereby formalized what has become known as the IT productivity paradox.

The paradox, of course, is that national accounts show trillions being poured into information technology but don't show a commensurate growth in productivity.

True believers, therefore, question the aggregate statistics or find other ways of arguing that the results do not reflect reality. Unfortunately it is simpler to believe, with Bishop Occam, that the numbers are right and that most boards, which usually rubber-stamp internally generated proposals that reflect the incentives to overhead growth characterizing systems deployment in most companies, simply make bad systems decisions.

In our example, board members need to consider two critical issues before making their decision:

  1. Will the business become more productive and make a better return on investment if it uses Windows or if it uses Unix?
  2. Does this decision affect the basic structure -- organization, compensation, controls -- of the business? If so, is one set of outcomes preferable to the other?

These are very difficult questions to which no complete answer can be given but we can start to address them by asking two much simpler questions:

  1. What is the relative productivity impact of Windows versus Unix in the enterprise?
  2. How does the Unix versus Windows decision affect the role and function of IT in the organization?

The productivity effect of failure reduction

Since the enterprise software selected, PeopleSoft, is client-independent we can focus here on the productivity effects of the two delivery mechanisms: Unix with smart displays or Windows client-server. We know Windows costs about $40 million more over the period, but how does it affect user productivity?

According to the Microsoft-sponsored NSTL Test Report: Microsoft Windows 2000 Professional Comparison of the Reliability of Desktop Operating Systems, Windows 2000 Professional has a MTTF of approximately 2,900 hours.

You should also, however, consult for a different result. Bugtoaster asks people to download and use a utility that reports the nature and cause of each failure the user encounters. The number of PCs reported running Windows 2000 SP2 increased from 1451 at noon on October 7th to 1556 at noon on the 15th. During those 192 hours the number of reported failures increased from 4915 to 6200. We don't know when the 105 individual PCs were added to the database but if all of them had been added at the beginning the numbers would reflect 1285 application crashes in 298,752 operating hours or one every 232.5 hours.

According to Microsoft's numbers the company can expect 13 desktop failures each day (the more realistic bugtoaster number is about 161 per shift) and one significant server failure every 15 days.

Each failure generates a crisis of some kind. Many result in partially completed transactions or loss of data. Each hurts user productivity both at the moment and in terms of the user's trust in, and so commitment to, the system.

One of the less appreciated, but more expensive, consequences of this type of daily failure coupled with continuous Windows product churn is that it becomes virtually impossible to separate systems support -- meaning Windows desktop, networking, and server support -- from application support.

A user hits "Ctrl-Shift-F3" and the PC draws an empty box instead of what's expected. What went wrong? Did the user load something that changed the PC's settings or libraries? Did one of the servers hiccup? Is this a transient network issue? Should the user have typed "Ctrl-Alt-F3"? There's usually no easy way to tell. As a result, first-line application support devolves to the Windows support staff. That has two negative, unavoidable consequences:

  1. Users learn to avoid experimenting with the system or the application and, as a result, they never learn to use either effectively.

  2. Support people quickly learn the application basics and start to help users figure out how to do things. Since they don't know the user's job, and so can't understand the application, this help is superficial and rapidly institutionalizes easily found, but ineffective, solutions to common problems that may be better addressed through procedures built into the software but obscured by the work-around.

  3. In contrast the Unix users will be unaffected by server failures and see, company wide, only about one smart display failure every two weeks. When a smart display does fail the user can continue to work on another device with no loss of data or other work and so these failures will not cause any kind of crisis.

    In the smart display environment there is no desktop operating system to support and so no support people to get between users and the application. That results in a very clear demarcation of responsibilities with systems delivering application processing, the application vendor responsible for functionality, and lead users responsible for teaching others how to use the application.

    Since these lead users know the job and are committed to the application, the direct effect is that more users learn to use more functions more effectively while the indirect effect is to build lasting trust relationships between the parties.

The cost consequences of these differences have never been formally measured, but they dwarf most other elements of the total cost of ownership. To have 5,000 users, the company would be about the size of PolyOne, a plastics and resins maker with about 9,000 total employees based in Cleveland, Ohio. This company generates annual revenues in the range of $3 billion. For a company this size, a 1 percent decrease in productivity due to user resistance and failure to properly use the application suite amounts to a $150 million revenue hit in five years.

If we assume a 30 percent contribution margin this means the company's operating earnings will be reduced by the direct cost of Windows ($64 million) plus the lost contribution margin (30 percent of $150 million) for a net reduction in operating earnings of $109 million.

Conversely, a 1 percent productivity increase due to effective use of the software, made possible by the absence of failures and the quality effects of peer-group-based, rather than systems-based application support, leaves the company with about $21 million in net gains in operating earnings (We get to this by subtracting 24 million in costs from the 30 percent contribution margin coming from the 150 million in productivity gains.)

In other words a 1% estimate of the productivity impact produces an estimate that the Unix decision is better, over the five years, not by the $40 million difference in costs, but by $130 million in operating earnings.

More realistic assumptions, such as those that include costs from Windows-induced downtime, produce larger estimates. For example, a 2.5 percent productivity variance produces an impact on operating earnings in the range of $750 million over five years.

The impact on the role of IT in the organization

A massive productivity benefit is one thing, but what are the implications for the business design? Will Unix use lead to a smoother, more efficiently functioning business? Does a Unix decision make a different contribution compared to a Windows decision?

From the board's perspective, organizational design is about aligning incentives with desired results. That's easy for areas like Manufacturing or Marketing where an obvious measure such as defect rates or sales relates naturally to growth in the power, prestige, and earnings of the executives involved. It is much more difficult, however, for the CIO role because the cost cutting mandate normally given Systems contradicts both its company wide service mandate and the normal incentives to budget and staffing growth felt by all executives

The organizationally right answer is to align the Systems department with the organization's revenue producers instead of with Finance, and then fund it from operations instead of from overhead.

This is easy in theory, hard in practice because it does not diminish the need to meet user expectations with respect to the most basic "automated clerking" functions that Systems undertakes. These, after all, were the functions that got Systems into the business to begin with and placed it, logically enough, as a cost center within Finance.

From an organizational design perspective what is needed is a simple way to make the CIO's ability to pursue his revenue side opportunities contingent on effective execution of the cost reduction mandate in the overhead side of the job.

That incentive structure comes naturally with Unix. Consider that the Unix CIO can move beyond service provision to revenue generation because he has two things going for him:

  1. he has his house in order;

    Job 1 in IS is always application processing -- but he's got that covered. To do it well and cheaply he's relying on very fundamental Unix advantages in scalability, reliability, ease of administration, and access to advanced software running on smart displays.

  2. and, his people have developed the strong relationships with users that are critical to partnering with them.

    Users know that the applications work and that systems people, who in true Unix style are encouraged to work directly with users, and are on their team.

For this CIO, the incentives are properly aligned. He is both personally and professionally better off if growth takes place at the revenue edge of the business, instead of in overhead. Cutting overhead increases user confidence and frees resources for use in revenue generation while increasing overhead creates diminishing returns to Systems as users lose confidence and growth resources are re-directed inward.

In contrast, the Windows guy gets user social approval by going along with perceived wisdom but must always disappoint because he can't deliver reliable application services. This isn't his fault, but the same social presures that cause users to approve of the Windows desktop preclude them from seeing it as the source of the problem. Unfortunately this conflicted user perception has some negative consequences: users may, for example, like the CIO as a person but refuse to partner with Systems people on new revenue generation because they don't trust Systems to meet their expectations.

As a result the Windows CIO can launch independent revenue initiatives, like a web site or selling homegrown software, but not merge his organization's efforts into those of the company as a whole. To do that he has to earn user trust - and that's something he can't do because but the technology and its organizational consequences effectively prevent it.

At the technology level each new Windows generation promises improved reliability and ease of administration but development is going in the wrong direction to actually achieve this. What he needs is simplification, but what he gets is increasing complexity. When Microsoft forces him to move from a 15 million line desktop operating system to a forty million line one he's being pushed in the wrong direction from the corporate viewpoint and the right direction from a personal one. To get reliability, and costs, under control he needs to reduce, not add, complexity - but the technology won't let him do that. He can't simplify his desktops without switching to smart displays and he can't get Unix like scalebility and reliability in his servers without switching them to Unix.

Fundamentally he can't do the job the company is paying him to do - the technology and its current direction simply won't let that happen. Instead, he's dragged along on an inexorable cost escalator as each new generation is more complex, more expensive, and more tightly integrated with other licensed products than the last one. What he can do, however, is grow his share of corporate overhead and increase his visibility in the executive suite -user unhappiness with his services coupled with the social acceptability of the Windows approach makes that both easy and rewarding.

In the real world you don't usually start out with a clear choice: Unix or Windows for 5,000 users who currently don't have systems support. Instead, you take over an existing infrastructure and plan for a long-term transition in which you re-vitalize your staff, develop trust relationships with users, and re-educate senior management.

It is extremely difficult to move a data center from a mixed or proprietary environment to Unix. That process, known as data center defenestration, is the subject of my book The Unix Guide to Defenestration and requires far more than technical change. The challenge is to change minds not just technology.

Using Unix enables, but does not ensure, success. Get people thinking in terms that are consistent with the core Unix ideas and you'll succeed; apply ideas and reflexes learnt in proprietary environments like Windows or OS/400 to Unix and you will soon have an expensive, dysfunctional, mess.

As a result the Unix decision allows the company to build in systems incentives that align with those of the organization as a whole, thereby enabling the CIO to join with his colleagues in earning revenues and building up shareholder value. The Windows decision, in contrast, permanently traps that CIO into spending monies earned by others while growing the systems department, and his personal career prospects, at shareholder expense.

We don't know of a reasonable methodology for estimating the impact this component of total cost of ownership has on companies. It is clearly very significant and probably more so, in fact, than anything else. Metaphorically, it amounts to the difference between operating the company as a sailboat dragging the Systems department behind it, and operating that same boat with the systems group functioning as an additional sail. We don't know how to measure it, but it probably spells the difference between success and failure for many companies.