|HOME||PUBLIC LIBRARY||ANOTHER PERSPECTIVE||INFOPERSPECTIVES||CONTACT|
When people talk about mainframe power, they usually mean the computing power the machines provide. But there's another kind of power that's big on big iron, and that's the power the machines use. All this power, thousands of watts, turns into heat. The heat means even more power consumption for cooling. To make things more complicated, there's a third factor: power density. As computer equipment becomes increasingly compact, it uses more power per square foot of floor space. So, unless IBM makes radical changes in its mainframe engineering, power and plant problems, already severe, can only get worse.
The last time IBM addressed the problems caused by high mainframe power consumptions was more than a decade ago, when it replaced bipolar circuitry with CMOS in its mainframe engines. At the time, CMOS was dramatically slower than bipolar, and IBM had to face a big sales challenge. It had to tell customers that CMOS was progress, even if it had engines with only a fraction of the power of older ones, even if it took seven or eight whole CMOS mainframes to do the work of one bipolar machine.
IBM promised improvements, and delivered, eventually. But it took IBM four generations of CMOS to not quite match the box power of the last of its big bipolar systems, the 1990 vintage 9021. When it came to system processing capacity, IBM didn't really leave the 9021 behind for eight years and five generations of CMOS.
Now all that is history, and even though IBM (or any other company making high end computers) could come up with bipolar machines that outran CMOS alternatives, nobody is going to turn back the clock. Bipolar technology inherently uses far more current than CMOS, current that is hard to supply and, when it is all turned into heat, as hard to cope with when it's going as it is when it's coming in.
Today's biggest IBM mainframe, 18,000 MIPS top of the z9 line, uses a lot of power, 18.3 KVA, but that's only half the current the smallest 22 MIPS 9021 used at the end of the bipolar era and an eighth the power the biggest 9021 models consumed.
Even so, at their most efficient IBM mainframes appear to use twice as much electric power per unit of processing power as X86 servers. You can see this very rough comparison if you compare the processing power of a heavy duty four chip (eight core) HP server, the DL 580 to a z9, as our table indicates. It's not possible to make a general comparison of the two machines, but if you divide X86 cycles by some factor, and we used 6 and 10 to get a range, the HP box comes out performing computing work that approximates a mainframe with somewhere from 1400 to 2500 MIPS of power. The z9 series ranges from 581 to 17,800 MIPS, but takes up a whole big cabinet or two, compared to the few inches or rack space the X86 machine occupies. The HP box we configured to use about 1 KVA, compared to the 6.8 to 18.3 KVA of a z9, had 16 GB main memory and 1.2 terabytes of hard disk space, a couple of gigabit Ethernet interfaces and four typical PCI cards for fiber, SCSI or other connections plus a pair of power supplies for reliability.
The X86 box, if you accept the back-of-the-envelope computing capacity calculation we've made, is not only twice as efficient as the biggest z9, it's 15 times as efficient as the smallest z9. This is because the z9, IBM's most power efficient mainframe when it's configured in large sizes, has tons of extra hardware in its cabinets when configured in smaller configurations. IBM puts in processing chips in blocks called books, groups of circuitry that can provide up to 10 working engines, and all the chips are kept warm all the time. Some of these chips may be locked out by microcode when a customer buys a system with less than the full capacity of the processor books in the box, but they use power anyway. IBM also builds machines with lots of memory even if it only turns on access to the amount of memory a customer pays for. Depending on how a z9 is configured, a customer may see 16 or 32 GB but have 64 or 128 GB of physical memory in the frame, all of it running whether or not it's being used for any customer purposes.
IBM undoubtedly has very good reasons for not actually turning off unused circuitry, such as the hot sparing that its big iron does automatically. But an outsider not privy to the secrets of IBM's mainframe engineering groups might well wonder why IBM hasn't figure out how to confine power to only what the customer uses and perhaps just a bit extra hardware here and there. After all, laptops do this to some extent, as do cell phones and, increasingly, servers built with so-called green criteria in mind.
So, while IBM has improved mainframe power efficiency with every generation and built its smaller mainframe, like the z890, with much less hardware (and power consumption) than its really big iron, as our chart [tim - please put in a link to the table I coded, which has been updated since I sent you the prelim by email so please use the latest version] shows, it has only made visible progress compared to itself, not to server industry as a whole.
And IBM's strategy has gone over well with customers, who can put faster, less power hungry mainframes in rooms that formerly housed older, less efficient models. But that comparison only makes IBM look good if users don't grow very fast. Customers who want to bring up new applications that demand serious increases in MIPS may well start thinking about electric power issues when their machines outgrow the limits of their physical plants. It costs a lot of money to power big computers, lots more to provide adequate battery and generator backup, even more to cool the glass houses and even more to juice the disks and other peripherals that inevitably grow when processors are upgraded.
It's bad enough paying the extra electric bills, which most computing departments can quietly slide out of their budgets into corporate physical plant allocations, but when that big iron forces a user to renovate its glass house or move to a new facility, a mainframe's high power requirements raise that recurring and perpetually ugly question: Is this machine really worth all the cost and hassle?
— Hesh Wiener December 2005