In a recent column, I said that the shift from a client/server architecture to Web services could open the door to Windows in data center. If Web services are mission critical, the infrastructure upon which they rely will have to be ready for the rigors of the data center. So too will the personnel assigned to the task of preparing those systems; and right now, the supply of qualified engineers needed to make Windows hum in the data center is more like a vacuum.
I asked Microsoft what it was doing to address those issues. As most of my readers know, I am loath to accept any vendor's "word for it." While I can't say for sure whether the next data center-ready version of Windows is indeed data center-ready, I will say that some of Microsoft's claims (such as benchmark results) were independently verifiable. The claims do demonstrate that Microsoft is serious about addressing the biggest concerns of data center managers. Can they succeed? That's a question only you can answer. Here's what the product managers have to say.
On the performance front, Microsoft product manager John Borozan lives and dies by the benchmark performance of Windows 2000 Datacenter Server. (I'll just call it WDS). He stresses the importance of two benchmarks in particular. The first is the Transaction Processing Council's TPC-C benchmark for on-line transaction processing. The second is SAP's Sales and Distribution benchmark, the results of which are certified by SAP.
"[WDS] may not have been a market success," said Borozan, "but we've been achieving our performance goals from a product success point of view." Borozan's evidence of this is what he calls a sneak peek at the performance improvements that will be a part of Windows .Net Datacenter Server (which I'll call WNDS). According to Borozan, "we've already incorporated those improvements into the Limited Edition of WDS, a version that moved us into sixth position in the single-server TPC-C rankings. We've never even been on that list and now we're up there against 128-processor systems with a 32-way box." Although it's not rated sixth anymore, the Windows-powered system is indeed on the list, currently ranked eighth in performance and third lowest in cost.
Borozan also points out that the same Limited Edition version of WDS is what has Unisys jockeying with IBM (using AIX) for the top spot in SAP's SD Benchmark. The current standings on SAP's site show Microsoft in the top spot on a variety of metrics including number of users, fully processed line items per hour, and number of dialog steps per hour. However, the SAP benchmark doesn't boil it down to cost. What's not clear is how the 100 servers (including 824 processors and 522 GB of RAM) that Unisys took to claim the overall title compare from a cost perspective to the 20 servers (including 480 processors and 520 GB RAM) that earned IBM the runner-up position. Also, it would be difficult to argue with anyone who claims the total cost of running 100 servers (including personnel, fault-tolerant configurations, etc.) might be dramatically higher than the cost of running 20 servers.
Tangentially connected to performance is scalability, and Borozan said that Microsoft isn't standing still there either. Borozan seemed genuinely excited about moving up the TPC-C list once Microsoft's work with its hardware partners starts producing 64-processor systems with 256 GB of memory. "Hopefully, we can challenge those big numbers where the 128-processor systems play," said Borozan.
In terms of plans to support both the AMD and Intel 64-bit architectures, Borozan reiterated that Microsoft intends to support both IA-64 and AMD's Opteron. When WNDS is released, however, it will only support IA-64. "Opteron," according to Borozan, "actually requires a different version of the operating system. We're not prepared to say when support for Opteron will be available, but software vendors should know that they won't have to recompile their applications in order to run on it." Despite Borozan's focus on performance, he agrees that most important criteria for WNDS' market success will be reliability. Before deferring to fellow WNDS product manager Josh Anderson, Borozan noted that WNDS, like WDS, will only be available through systems manufacturers and that those system manufacturers have to live up to certain standards.
"For example," said Borozan, "every single OEM that offers Datacenter has to guarantee 99.9 percent (three nines) uptime." This is just two nines short of the theoretical 99.999 percent availability limit. In order to sell WDS and WNDS, OEMs must offer 24x7 support that's mutually staffed by the OEM and Microsoft via a single queue. Also, they can only offer configurations that are fully tested right down to the printer drivers. Borozan claims that this mimics the Unix model that customers have indicated they like.
Regarding reliability improvements from WDC to WNDC, Anderson will focus on reducing both planned and unplanned downtime. This focus, no doubt, is inspired by Windows' reputation for requiring more frequent reboots than its competitors. Also, Anderson claims that the maximum number of nodes in a Windows cluster will be doubled from four to eight. His claim reminded me of Oracle CEO Larry Ellison's assertion that Microsoft and IBM's non-mainframe systems take a "shared-nothing" approach to clustering in which reliability may actually decrease as more nodes are added. In a share-nothing approach, according to Ellison, the more nodes that are added to a cluster, the greater the likelihood that one of them will go down, rendering the entire cluster useless. I plan to address that claim in an upcoming column.
Also making an appearance in WNDS are the HOT RAM and HOT PCI features that allow memory and PCI-boards to be added and removed without bringing the server down. "This will help reduce the amount of planned downtime," said Anderson. To address one cause of unplanned downtime, Anderson said, "We've also improved Internet Information Server so that one application can't bring another one down. Now, if one Web application goes down, you can recycle that process and relaunch it without affecting the other Web applications running on the same system."
A shortage of qualified experts
All the improvements in the world to a data center-ready operating system will be a waste of Microsoft's time unless qualified people to deploy it and keep it running can be found. A typical Microsoft Certified Systems Engineer will not be able to help you in data center situations. Microsoft acknowledges a shortage of qualified experts in the area, and it's product manager Jerry Smith's job to fix that problem. According to Smith, "We hear the MSCE problem loud and clear."
Smith personally authored the curriculum that he also administers to those seeking training on WDC. Since October 2001, about 360 people have passed through his six-day course; and to date, Microsoft hasn't charged anybody. Unfortunately, Microsoft has yet to offer a special data center-level of certification. But now that he's worked the kinks out of the curriculum, Smith has some changes planned, including some sort of special certification.
For starters, he needs to clone himself. Smith is working with WDS OEMs Fujitsu/Siemens, IBM, and Unisys to roll out their own training programs. The first of these will be run by Unisys next month. Later this year, IBM will roll out similar sections of the course at its training facilities in Europe. Microsoft will continue to run courses out of Redmond on a monthly basis but those courses will no longer be free. Smith expects them to run in the $1,500-$2,000 range and information on them is available on the Web.