Ten into one might go
Sometimes the maths is so simple you don’t even need your toes: the fingers of both hands were quite enough when I spent a day recently chatting with some guys from Intel. These chaps wanted me – and, by extension, you my readers – to understand that Itanium, and the whole concept of 64-bit computing, is alive and well. They’d brought some other guys along with them who told us all what to expect from a full 64-bit platform. It seems that what’s going to save Intel’s Itanium range of processors is server consolidation and virtualisation. Getting on the consolidation bandwagon produces an irresistible impetus toward bigger and bigger operations, and it’s becoming almost inescapable in the corporate sector. In short, Intel wants us to believe we can get one Itanium to do the job of ten old servers.
They’d brought along a lot of graphs and charts to support this assertion, including some baseline numbers about overclocking. It seems that most of the time an overclocked CPU draws 100% extra power in order to deliver only 20% of extra processing power. However, you can underclock a dual-core CPU by that same 20% and still get from it 178% of the throughput, because there are two nice lightweight cores sharing the workload. This is excellent news for Intel, because it puts the skids under those online communities who rave on about overclocking as a hobby/way of life. This fact underpins Intel’s whole philosophical roadmap for Itanium machines as the decade unfolds; namely, “the multicore Itaniums are coming”.
There was even a guy from Microsoft along at the meeting bearing more charts that depicted what software would be available to run on them, and according to him they’re expecting the main take-up of 64-bit multicore versions of Windows Server to be in the “eight-way or greater” server market. That means putting eight of these proposed twin-core Itaniums, or 16 distinct CPU cores, into a single server box. This brought a flashback to my mind, to a smallish, suitcase-sized object that an IBM demonstrator placed on a pedestal at LotusSphere a couple of years ago, but then couldn’t demo because the convention centre couldn’t supply enough three-phase power to fire up all its power supplies (even when running in a cut-down mode it made a pretty effective hairdryer). The scale of support operation needed to be able to cope with boxes containing that kind of processor density is going to be rather larger, I suspect, than most of us are used to paying for.
With that class of server in mind, it’s no wonder both the Intel and Microsoft guys believe the benefit to be expected from server consolidation will be a ten-to-one reduction in the overall number of servers a corporation needs to deploy. This is an amusing little number: for one thing, it wasn’t that long ago when Microsoft was quoting one-to-ten as the ratio of servers to average number of users across every type of business. Now, suddenly, we’re able to opt for compressing by the same factor again, saving space by buying ever-chunkier boxes of server hardware. But make no mistake, this is also an attractive number. The largest server room I’ve ever frequented had 13 aisles, with around 20 6ft-tall rack-mount cabinets per aisle. If that consolidation factor really were achievable, those guys could reduce down – allowing for the usual maintenance churn and some extra gain by fiddling about with virtualisation of test and live systems – to just a two-aisle operation.
However, here’s where the count-on-your-fingers maths cuts in. I can imagine some situations in which that ten-to-one factor might prove to be a light estimate. Big corporations have for a long time been notorious for buying the very top-of-the-line servers to support their annual budget-inflation habit, but then chickening out over equipping them with more than one processor because their internal technical standards have never been revised since the days when multiprocessor operation was a novel (and occasionally dangerous) concept. I hear many system design and server sales people having a good old chuckle about Cassidy’s Crazy Notion on this subject, but I’d urge them all to take a look at the servers listed for sale on Ebay and to focus their attention particularly on the clearly ex-corporate machines. How many do they see that actually have no room for more processors, compared to those that have been run for a half-decade with only a quarter of their potential processor slots filled?