Two years ago, I started a research project that sought to examine energy usage patterns in consolidated, scale-up servers vs. distributed server farms. My theory was that enterprises that were deploying distributed, towered servers—with utilization rates often less than 20 percent per server—were probably wasting a significant amount of energy. To compare energy use in each environment, I chose to concentrate on scale-out, x86-based towered servers and scale-up mainframes. The data showing energy usage characteristics of x86- based servers was extremely easy to obtain. Loads of x86-based server energy usage data can readily be found on various Websites.
Finding data on mainframe energy usage, however, was far more complex. To measure energy consumption on a mainframe, factors such as the kind of workload being processed and system utilization rates came into play. At the time I started my research project, the best data I could obtain on mainframe energy usage essentially consisted of IBM guesstimates (educated guesses) about what the maximum power consumption characteristics would be under certain workloads at certain utilization rates. I don’t do reports based on guesstimates, and because no solid, concrete data was available, I killed the project.
In October 2007, IBM announced its mainframe “gas gauge,” a systems monitoring tool that can generate precise mainframe energy usage reports. Using internal sensors, IBM’s gas gauge can provide data on kilowatts used, BTUs generated—and even CPU utilization characteristics. This data can be viewed on IBM’s System Activity Display, and also can be fed into usage reports for ongoing analysis and resource utilization planning purposes. Additionally, when used in conjunction with IBM’s Power Estimator, enterprises can estimate energy use and cooling requirements—and tactically and strategically manage systems resources accordingly. The combination of the gas gauge monitoring tool with IBM’s graphically driven management activity display and related performance analysis/control tools enables mainframe users (and research analysts) to obtain an accurate picture of real-world mainframe energy usage requirements. Using gas gauge data, IBM has derived that a single mainframe processor with z/VM virtualization can perform the work of tens to hundreds of x86 processors using only 2 to 10 percent of the energy used by towered, underutilized x86-based servers!
Reasons for Efficiency
The primary reason mainframes are so efficient when compared to underutilized x86-based servers has to do with the level of sophistication of the mainframe virtualization capabilities. Mainframe resources have been virtualized for almost 40 years, and over this time, hardware and software developers have automated memory sharing and management schema, created advanced CPU resource sharing designs, and have continually improved systems scalability through “extreme virtualization.”
Another reason mainframes shine when compared to x86 architectures also has to do with the centralized design of the mainframe. The distributed nature of scale-out architecture forced a lot of Network Interface Cards (NICs) and external hubs/switches/routers to be used—adding even more energy inefficiency to scale-out architectural designs. These devices not only add to the energy burn, but also create additional cooling costs. Combine these devices with distributed server energy drain and it boggles the mind how much energy might be needed to cool tens, hundreds, or thousands of distributed servers that run at only 20 percent capacity. A third reason mainframes outshine distributed x86- based servers deals with processor efficiency. A single processor chip is capable of efficiently executing hundreds of workloads. For example, a single mainframe processor running z/VM virtualization can perform the work of tens to hundreds of x86 processors because its processor speed now exceeds 4 GHz—and because these chips have been designed to run many mixed workloads at high utilization rates.
Side-by-Side Comparative Data
Currently, IT buyers have solid comparative data that shows scale-up systems are far more energy efficient that scale-out designs. But what is missing are side-by-side comparative standards. According to David Anderson, IBM System z green evangelist, standard measurement criteria are now evolving to help IT executives compare energy usage by system type. The Green Grid standards organization now measures energy use/system effectiveness using Power Usage Effectiveness rating metrics. As standards evolve, “comparing server types side by side will be far simpler” he observes.
Meanwhile, using its gas gauge monitor, IBM has been aggressively collecting mainframe energy usage data for the past nine months. In October, IBM announced its initial results as part of its gas gauge press announcement (these results showed the mainframe using 2 to 10 percent of the energy consumed by hundreds of x86-based servers). Clabby Analytics expects the next big gas gauge announcement will come when IBM has gathered more, real-world, empirical data on its new System z10. It’s reasonable to expect IBM to publish z10 energy consumption data after a few more months of data collection. Advice to It Executives IBM’s mainframe gas gauge shows huge energy savings can be realized using mainframe architecture as compared with distributed systems architecture. And with faster processors on its newly announced z10 Enterprise Class, IBM has been able to greatly increase mainframe processing capacity (by 70 percent), further widening the gap between mainframe architectures and distributed server farms.
But what should your organization do with this information? Anderson points out that the gas gauge, when combined with other IBM management products and tools, enables IT executives to “use the gathered energy information to make informed tactical and strategic decisions based on energy usage.” According to Anderson, the new goal IT executives are working toward is to “improve performance using the same or fewer watts”—or rephrased as “do more work using less power.”
As for my advice from a research analyst perspective, I suggest that:
• If your organization runs Java/Linux-based workloads on x86-based distributed computers, it should evaluate moving those workloads to mainframe environments. Most IT executives I’ve spoken to don’t know that IBM has special pricing in place for IT shops that want to deploy Java/Linux workloads on mainframes (running these workloads on IBM’s Integrated Linux Facility [IFL] significantly lowers both hardware and software costs). Compare the costs of your x86-based hardware, software licensing, security, management, and high-availability (backup systems) to mainframe costs—and then throw in energy savings costs—and your organization will likely find mainframes are far less costly to operate than x86- based server farm environments.
• Mainframes currently run z/OS, z/VM, and Linux operating systems. Sun’s OpenSolaris is currently being tested as a guest operating environment, so mainframes soon will be able to run Unix workloads, too. But mainframes don’t run Microsoft Windows, nor are they likely to in most of our lifetimes. If you’re running Windows applications, consider moving them off energy-hog distributed systems and onto blade servers. Blades, like mainframes, save tremendous amounts of energy by reducing networking component and energy consumption costs, by relying on fewer and more efficient power supplies for internal servers, and by running at high utilization rates.
• If you can’t eliminate your x86-based server farm, for whatever reason, you must make an effort to improve the utilization rates of the equipment you now own. Consolidate and virtualize wherever possible within your IT environment to reduce energy waste; to lower acquisition costs (because you’ll get better utilization out of servers that you already own rather than having to purchase new ones); and to reduce software, highavailability, and management costs.
• The energy savings a scale-up mainframe can deliver when compared to underutilized scale-out servers are tremendous. But it’s important to note when considering mainframe technology, to ensure maximum payback on a mainframe investment, IT executives must ensure their respective enterprises have enough work to process to continually drive a mainframe at 80 percent-plus utilization to achieve the best return on investment in this type of architecture. Otherwise, a different scale-up architecture (such as an optimized, consolidated, virtualized Unix server) may make more sense.
Finally, IT executives need to look beyond the word “mainframe” to what’s really happening here. What IBM has implemented is a monitoring/measurement system that clearly shows how much energy can be saved by using scaleup system designs. But IBM’s mainframe gas gauge isn’t the main point; the main point is that we now have empirical data that shows how efficient scale-up designs really are (or conversely, how stupendously wasteful scale-out tower and potentially rack designs really are). This knowledge is extremely important to IT executives looking to reduce operational costs while at the same time helping to save our planet by reducing energy waste.