Power Systems Memory Prices Slashed to Promote Virtualization
August 11, 2008 Timothy Prickett Morgan
While server makers are, generally speaking, pretty excited about server virtualization, their excitement is not just limited to the idea that the advent of sophisticated virtualization hypervisors and the efficiencies that consolidation affords compels customers to spend money on new gear when they might otherwise not spend anything at all. Server virtualization and the consolidation it allows also compels customers to buy beefier machines, ones with lots of processor, memory, and disk capacity, and I/O bandwidth. These are machines that are inherently more profitable than the little boxes sprawled all over the data center these days.
Since the advent of enterprise-class logical partitioning with the AS/400 back in February 1999–yes, it was commercially available years ahead of the partitioning from Unix vendors and only a year after VMware was founded–IBM has been talking about the economic, operational, and technical benefits of server consolidation, much as the advent of (relatively) cheap telecommunications in the early 1990s engendered a wave of data center consolidations. In the wake of the minicomputer and PC revolutions in the 1970s and 1980s, companies spread their computing all over their locations, even large ones, and that created a lot of operational and capital costs. So, too, does having servers sprawled within a single data center and not efficiently using server resources. But server virtualization is difficult, which is why you could say that VMware took until 2004 to field an enterprise-class product, and you could argue that this didn’t really happen with ESX Server until 2007. And it is why Microsoft has struggled to get out Hyper-V even this year.
Even though server hypervisors are getting plenty of help from the electronics in their Power, Itanium, X64, and Sparc processors to make them more efficient, no hypervisor runs without some kind of performance penalty and, equally importantly, consolidating many smaller server workloads down onto one larger physical server (and thereby driving CPU usage from 5, 10, or 15 percent to maybe 60, 70, or 80 percent on average) means having to add a lot more memory to the boxes than you might think. How much more depends on the nature of the workloads, the hypervisor in question, the iron, and other factors, and no server or hypervisor maker is explicitly telling customers what the memory overhead is. But the fact that IBM last week chopped memory prices on its Power Systems Power6-based servers as well as on a number of its predecessor System p and System i machinery that uses the same memory cards, is a pretty good indication that there is a penalty and IBM wants to help cushion the blow in an effort to promote the use of its PowerVM hypervisor for workload consolidation.
Such a price cut may seem counterintuitive, particularly in a cut-throat server market with skinny margins on all but the biggest iron. “People are buying more processors, and our average unit revenues for 520, 550, 570, and 595 machines are going up, and memory is a big component of this rising average,” explains Scott Handy, vice president of marketing and strategy for IBM’s Power Systems division. This is the kind of trend they give you a bonus and stock options for, or at least I thought. But, as it turns out, Handy wants to accelerate the adoption of the PowerVM hypervisor–formerly known as Virtualization Engine generically and on the System i line and Advanced Power Virtualization on the System p line–and to do so, he took a red pen to the prices of memory used on these Power-based servers.
You can see the price cuts in this table I created, which explains what features have had their prices changed and by how much. Memory features on Power Systems sometimes are just plain old DIMMs (such as in entry machines) or are memory cards with memory capacity turned off plus memory activation fees (in plain English, there are two pieces to the memory price instead of one), so characterizing the price cut is a bit difficult. Memory feature prices were cut on these four classes of machine (and also on the Power 575 supercomputing node) by 26 percent to 50 percent, but the average across all features works out to 30 percent. IBM is happy with system-level prices on Power 520 and Power 595 machines, so even as raw memory feature prices were cut, Power 520 system prices and Power 595 memory cards with no capacity activated were raised by an equivalent amount to keep the overall Express configuration prices for these machines the same. On the Power 550 machines, the memory price cuts are being passed straight through, and the base Express configurations are seeing their prices drop by 14 percent to 15 percent, while Power 570 Express configurations will see a 7 percent drop.
“The intent of these price changes are to favor customers who do virtualization,” says Handy–plain and simple. But of course, nothing is ever really that simple.
By dropping main memory prices on the Power Systems line and on older System i and System p machines (only a few of these machines, mind you), any customer who has a workload that can benefit from having more main memory thrown at it will also benefit. Modern, interpreted languages such as Java and PHP love main memory as much as they love CPU cycles and cache memory, for instance. And in many cases, traditional online transaction processing workloads written in RPG and COBOL can see their performance boosted as much by adding main memory in big gobs as they see from having more processors underneath them. (This has been true since the first days of the AS/400 line, and Big Blue used to hate it when I could show when a memory upgrade costing significantly less than a processor upgrade yielded the same performance boost.)
No matter what, the price decrease on memory features and capacity is good for all Power Systems customers in the short run, regardless of their plans to adopt virtualization.