When Linux was ported to IBM’s venerable System/390 mainframe hardware just over a year ago, something magical happened: companies considering throwing out their mainframes started to re-examine these plans. Then IBM announced support for Linux as an integrating, unifying platform on its entire line of e-business servers, and suddenly big iron was back in vogue. The wild and wooly world of Linux geekdom got a large dose of pinstripe suit boardroom acceptance, and the hardware fashion mavens started saying, “Retro is so chic this year!” instead of, “A mainframe. How…quaint.” In short, the marriage has been a happy one so far.
A recent IBM press release highlights two reports by industry analyst groups that show substantial cost savings for Linux on big iron as compared to horizontally-scaled rack-mount server farms. Especially in the energy-starved Silicon Valley area, IBM says, the power consumption alone can cost justify deployment of a single mainframe to replace hundreds of rack servers. A report by Hurwitz Group, based in Framingham, Mass., for instance, forecasts a savings of nearly US$200,000 annually if 500 Sun UE250 servers are replaced with one of IBM’s new zSeries mainframes. These numbers are certainly enough to grab attention. The underlying issues, though, are complex, and the raw numbers don’t tell the whole story.
Mainframe hardware is not cheap, and neither is the software that controls it. While Linux itself is free or, in a commercial package, nearly so, the underlying VM operating system that allows hundreds or thousands of Linux instances to coexist is considered a costly line item even in mainframe circles. Administering System/390 Linux itself is very similar to administering Linux on other platforms, but getting the underlying mainframe systems up and running requires some specific talent that might not exist in an otherwise-Linux-only shop.
On the plus side, though, the mainframe rewards its users with reliability that redefines the word “exceptional.” Mean-time between failures (MTBF) is calculated not in hours but in decades. If downtime for an e-commerce business means thousands of dollars per minute in lost revenue, then avoiding even a few hours a year of service interruption can pay for a lot of hardware and software. That “costly” license for VM, used as a hypervisor, doesn’t look so bad when it can be amortized over five thousand (or more) individual ASP hosting customers. As with so many things, the economics of mainframe Linux are heavily dependent on the business needs of the enterprise that deploys it.
Looking at the specific issue of energy usage, for instance, the basic numbers seem simple enough. Published ratings for a Sun UE250, according to The Hurwitz Group, place its consumption at 360 watts. The corresponding figure for an IBM zSeries is 5.3 kilowatts. Hurwitz adds a 30% overhead to each of these in their calculations, to account for environmental consumption by building air conditioning equipment and lighting. Hurwitz Group’s calculations assume that one zSeries can replace 500 of the Sun machines in a typical managed hosting environment. Based on these assumptions, and an electricity cost of ten cents per kilowatt-hour, Hurwitz Group calculates a daily cost of $561 for the Sun machines and $32 for the zSeries. That becomes a difference of $193,085 over the course of a year.
Do the Numbers Add Up?
The most easily questioned number in these calculations is the assumption that one zSeries equals five hundred Sun SPARCs. Is that realistic? Rich Ptak of The Hurwitz Group says it is. “We tried to do an equivalency,” he explains. “We talked to some system integrators…with experience in both kinds of systems. We took an extremely conservative approach; we didn’t specifically tune the mainframe environment.” Recent customer profiles published by IBM and others seem to suggest that the figure of 500 is actually somewhat conservative. Telia, a European telecomm company, replaced over 700 Sun machines with one zSeries recently.
IBM’s Pete McCaffrey, program director for Enterprise Server Marketing, concedes that the raw MIPS of a mainframe do not equal the raw MIPS of five hundred Sun servers. But he says it’s real-world throughput, not published MIPS, that should be compared, and that segregation of total capacity onto hundreds of individual servers prevents the kind of dynamic load sharing that a single mainframe can do. “What we find in a customer environment,” he says, “is that a lot of times those Sun processors go under-utilized. Utilization rates are often only twenty to thirty percent of capacity.” The zSeries, he claims, “drives the application to a higher utilization rate.” The mainframe takes advantage of the bursty nature of system load in e-commerce and web applications, just as a utility company takes advantage of load averaging among thousands of consumers, and thereby avoids the need for a system that can meet the peak demand of all of the Linux instances at the same time.
Still, McCaffrey concedes–as have other IBM sources–that the zSeries is not a good fit in certain CPU-intensive applications. Its performance strength lies not in MIPS but in I/O bandwidth. “It is application-dependent sometimes,” explains McCaffrey. “In scientific applications, for example, you won’t get the same equivalency.” But he says that business deployments at customer sites are bearing out IBM’s claims of zSeries benefits. “As ASPs grow and build their customer base,” McCaffrey asserts, “dedicated [rack mount] servers can become an expensive proposition.” An ASP hosting company that is adding twenty rack mount server per week, McCaffrey observes, rapidly becomes mired in bad economics.
From The Matterhorn Group’s Perspective
The Matterhorn Group, based in New York, approached the question of mainframe economics somewhat differently but arrived at some of the same conclusions as The Hurwitz Group. They found the power savings are indeed real, and that the break-even point in terms (only) of electricity consumption is about thirty Linux instances on the zSeries. Furthermore, Matterhorn Group’s report cites floor space savings as another significant cost reduction.
But the Matterhorn Group report also raises some counterpoints to this argument, again reinforcing the idea that these economic analyses are not trivial. The baseline platform cost in the mainframe world is larger, meaning a customer needs to be running many Linux servers right away in order to justify the initial investment. Furthermore, the performance calculations may or may not favor the zSeries, depending on the application. Mainframe disk storage arrays are designed with a different set of underlying assumptions about the needs of the environment, meaning that system integrators need to think about storage planning in a somewhat different mindset than with horizontally-scaled server farms. Finally, the use of VM as a hypervisor to support those hundreds or thousands of Linux instances means that the installation will need at least one person with a mainframe-oriented skill set.
Another, subtler, factor cited by The Matterhorn Group is the psychology of shared server environments. Even though the VM hypervisor has a decades-long track record of superb security and inter-application isolation, there is a reluctance on the part of some hosting customers to share the hardware with anyone else. Paranoia runs deep in any good system administrator, and those who purvey zSeries application or web hosting have to overcome their customers’ instinctive distrust of shared environments. If ever an operating system could be that secure, VM is probably the one. It’s simply a question of comfort level for the customers, and only time will tell how that question will be answered.
The Matterhorn Group’s report still concludes that the zSeries mainframe can be “really exciting” as an option for providers. Especially given IBM’s new dynamic licensing model, which allows customers to pay for mainframe middleware on an actual-usage basis rather than system capacity, a larger installation could realize significant savings.
Tomorrow, in Part 2, Sun responds and author Scott Courtney does an in-depth comparision of today’ the two “glass house” platforms–a mainframe running Linux or a Sun server farm running Solaris. //
Scott Courtney is technical editor at internet.com.