In a press release dated January 23/02 Red Hat claimed:
"IDC Research reveals a 45-80 percent lower total cost of ownership (TCO) for Linux on the Intel architecture over RISC/Unix environments."
Since Linux is Unix, I read this as "Unix is 80% cheaper than Unix" which didn't seem to make a lot of sense. So I decided to investigate.
The actual whitepaper:
is an interesting piece of work.
Their methodology seems to have been to select two groups of users:
This approach sounds sensible and certainly met their sponsor's requirements but doesn't provide a fair comparison because the populations from which the samples are drawn aren't remotely comparable. The Linux users they queried typically deployed Linux in small, limited purpose applications using brand new gear while the RISC population typically used older, larger, gear, running multiple concurrent applications under the control of more traditional management.
In effect the sampling choices seem to have rigged the study to produce data in which:
As a result IDC found that the cost of running sendmail on Linux for about 1,000 users on a new PC is much less than that of running Lotus for 7,000 users on an HP N-Class - that's also running Oracle, a half dozen custom applications, an intranet, and the local single-sign on system.
To do this kind of comparison fairly you need to work through TCO in terms of realistic business scenarios - which is just what I've been doing in my series on TCO for Linuxworld.com.
Suppose, however we restrict ourselves to just correcting the obvious problems with the IDC report - i.e. compare Linux on Intel to Unix on RISC while using current costs for both sides and trying to eliminate the impact of management and related issues? To do that, we can assume that both kinds of gear are run by the same people, using the same applications, and then restrict ourselves to comparing the capital and maintenance costs of the two kinds of Unix system.
We can do that at three levels:
A 1400C with two 933MHZ P3 processors, 512MB of RAM, and two 36.4GB SCSI disks can be had, with Linux pre-loaded, for about $2,950 from Dell. As a general purpose small office server running applications like e-communications, web services, or, file and print sharing, this combination is a clear winner.
It is possible to offer a RISC competitor for it, but the comparison is to a strawman rather than something you'd expect people to do. The numbers canbe made to work out in favor of the RISC solution too: you can run the same software for the same number of users with a Sun 100 workstation ($995) by adding disk (20GB for $300) and memory (1GB for $790) for a total of about $2,085; but the performance hit makes it a specious comparison - nobody sane is going to make this choice.
Get a little larger and the comparisons get more realistic. For example, Sun's successor to the 450 (originally touted as its Compaq Killer) is the V880. On the numbers this machine narrowly beats the Proliant ML750 and its cousins throughout their range:
| Scale | Sun Configuration | Compaq Configuration | Comparability issues | RISC Capital Cost | Compaq Capital Cost | RISC three year 4-Hour on-Site Support | Compaq three year 4-Hour on-Site Support | Three Year RISC Advantage | RISC Advantage % |
|---|---|---|---|---|---|---|---|---|---|
| Small | V880, 2 x 750MHZ, 8GB, 6 x 36.4GB FC-AL Disk, DVD, tower | ML750,2 x 700MHZ, 8GB, 6 x 36.4GB SCSI Disk, DVD, basic rack |
|
$29,995 | $34,152 | $13,320 | $12,575 | $3,412 | 7.9% |
| Medium | V880, 4 x 750MHZ, 8GB, 6 x 36.4GB FC-AL Disk, DVD, tower | ML750,4 x 700MHZ, 8GB, 6 x 36.4GB SCSI Disk, DVD, basic rack | $49,995 | $55,360 | $13,320 | $12,575 | $4,620 | 7.3% | |
| Large | V880, 8 x 750MHZ, 8GB, 6 x 36.4GB FC-AL Disk, DVD, tower | ML750,8 x 900MHZ, 8GB, 6 x 36.4GB SCSI Disk, DVD, basic rack | $102,655 | $112,390 | $13,320 | $12,575 | $8,990 | 7.8% | |
| All pricing is from the Sun and Compaq websites as of Jan 27/02. | |||||||||
Of course if we try to go further upscale the wheels fall of the comparison again; even the four processor Proliant stretches Linux about as far as it goes today. You can, in theory at least, use a network of linked Linux machines to solve some of the same problems a larger RISC box is designed to deal with --but it's no more practical in reality than comparing a Sun 100 running Solaris to a Dell 1400 with FreeBSD. You can do it, but it doesn't make sense.
We could, for example, pretend to believe that all transactions are fully independent and so compare a large RDBMS instance running on one Sun 6800 to something similar on an IBM Xseries with 128 Myrinet boxes running Linux. With only a minor cheat (using Sybase on the 6800 and Oracle on the IBM) the five year TCO for the RISC solution cane made to come out more than 30% better. Unfortunately, this isn't realistic: fully independent transactions are found in benchmarks, not businesses.
So does this line of argument prove that RISC is cheaper than Intel/Linux? Of course not, these comparisons are guilty of the same fault that undermines the IDC report - I'm implying functional comparability where none exists.
In the real world it doesn't make sense to compare the costs for a ninties N-Class bought from HP by some ex mainframer who takes eight months to get it implemented and thinks it perfectly reasonable to have teams of people writing reports about how well they're going to serve user needs next year with costs applicable to a new Dell mini-server whose recipient orders it with Linux preloaded and puts it into production by plugging it in.
That's what's wrong with my comparisons too: I'm implying the functional equivelance of each pair of systems:
Even the most realistic of them, those between the Proliant and V880 configurations, have problems. If the smaller Proliants would meet processor demand, most people would agree that running Linux on a two processor server from Dell would make more sense - and save about $20K in the process. If, on the other, you need an eight processor box like the high end Proliant, you are going to want to run Unixware or Solaris on it, not Linux.
So what kind of comparison does make sense?
Lets recognize that Linux, BSD, Darwin, Solaris, Tru64, and Unixware are all Unix variants. The right comparison isn't between Unix and Linux; it's between all of Unix, including Linux, and the range of Microsoft Windows products.
On that basis the bare bones cost comparison is a no brainer: Unix wins easily. The issue there isn't whether Unix is cheaper -that's a gimme- the issue is whether it's smarter and that depends mainly on what you do with it. Is Linux on Intel smarter than Windows on Intel? For most Linux applications that's a comparison between a cheap and highly reliable system on one side and Windows on the other - not exactly a difficult choice. Is Unix on SPARC a smarter choice than Windows Advanced Server on a couple of rackmounts of little machines? That's the same question, with the same answer, just on a different scale.
Getting people to make smarter choices means getting them to make informed choices. That's hard, most people are deeply resistant to new information and eager to grab onto anything, however dishonest, that supports not changing their opinions. That, more than anything else, is what's wrong with this IDC report - that Red Hat would sponsor something so distructive of the Linux/Unix community it relies on for commercial success.
Click here for access to the Linuxworld.com TCO series. Click here for access to book site for The Unix Guide to Defenestration.
Want to send the author a comment?click here