Datacenter: Up Close and Personal

Playtime in a Unisys test lab

When I interviewed several Windows 2000 Datacenter Server administrators for "Datacenter in Action" (http://www.win2000mag.com, InstantDoc ID 20777), they each made running their systems sound easy. Too easy, I suspected—until I tried a Datacenter system. A big advantage of a Datacenter system is that the system arrives with the OS installed and tailored to your needs. The OEM sets up the tools, customizes the snap-ins, and writes additional custom utilities—an administrator's dream. And as complicated and powerful as Data-center might be, performing typical administrative tasks is straightforward and easy.

Hardware
One rule that Microsoft sets for Datacenter OEMs is that the OEMs' technical personnel extensively test each computer before delivery. I visited Unisys's Datacenter server test lab and, of course, chose to play with the biggest and brawniest server there: the Unisys e-@ction Enterprise Server ES7000. This refrigerator-sized system uses Unisys's Cellular MultiProcessing (CMP) architecture and is a marvelous specimen of computer engineering. Unisys licenses the CMP design to other Datacenter OEMs.

What you call a backplane in a PC is a midplane in this behemoth (the device sits in the middle of the box). Components such as processors and buses attach to both the front and the back of the plane. Memory also resides on it. You can access the midplane through the box's front and back doors. Each side of the midplane can have a separate power supply.

The ES7000 that I worked with contained 32 Intel Xeon processors; the machine is ready to receive Intel's 64-bit Itanium processors when they become available. You can mix and match processor types within the machine but not within individual partitions. The ES7000 arranges processors in eight groups, called sub-pods, of four processors. Each sub-pod connects to a Level 3 cache module (the processors provide Level 2 cache). You can assign sub-pods to any partition, although this option restricts you to assigning processors in groups of four.

Each sub-pod has a direct I/O bridge (DIB) that has three PCI buses holding four PCI slots each. Do the math: That's 96 PCI slots. Four memory modules, which Unisys calls Memory Storage Units (MSUs), hold 16GB of memory each. These MSUs look like SIMMs on steroids. The Level 3 cache modules provide data to the system at about 6 times the speed that system memory does. Unisys engineers told me that for day-to-day operations, most applications will find necessary data waiting in the cache 95 percent of the time.

A device that Unisys calls a crossbar uses point-to-point connections between memory, processors, and I/O components. The crossbar is the core of Unisys's CMP architecture, boosting the ES7000's performance because data doesn't need to cross the bus to move between these components.

Cooling this box isn't a job for a standard fan. The ES7000's three impellers look like small jet-engine fans. "Almost everything in the box is hot-swappable," said a lab technician, as he yanked an impeller out of its fitting. Immediately, the noise level rose as the remaining units increased their speeds to compensate. When the technician replaced the device, the other impellers slowed their speeds. Although hot swapping isn't a new concept, per se, I'm a Windows administrator, so I typically don't see much of it. This demonstration of the ES7000's hot-swapping capabilities impressed me.

Managing Datacenter
Datacenter servers offer several management tools, some from Microsoft and some from the OEM. OEMs typically load Win2K Server Terminal Services in Remote Administration mode on the Datacenter server. Administrators can then use Terminal Services Client to access the Datacenter server from their Win2K Professional machines.

Datacenter's primary management tool is Microsoft's Process Control, which administrators can use to organize and manage processes and the resources processes utilize. Process Control runs as a service on Datacenter, and administrators can use either a Microsoft Management Console (MMC) snap-in or a command-line interface to manipulate Process Control's configurations and functions. (If you install Process Control directly on Win2K Pro, you can also remotely administer Datacenter processes and resources without using Terminal Services.) You can use Process Control to set limits on a process's use of virtual memory or assign processors to a process or group of processes.

Datacenter uses job objects, new kernel objects that Microsoft introduced in Win2K. Job objects control groups of related processes, letting you manage the processes as single units. You can use Process Control to set rules for how much of the system's resources a job object­controlled process group can utilize. (For more information about job objects and Datacenter hardware, see Greg Todd, "Win2K Datacenter Server," December 2000.)

To monitor processes, Process Control launches a separate tool, procconmd8.exe. This program is the enforcer (or as Microsoft more mildly terms, the mediator) that polices your system for broken configuration rules. Even if you stop Process Control (e.g., to change configuration schemes or rules), the mediator continues running, safeguarding the environment.

Unisys, like most Datacenter OEMs, supplies proprietary tools in addition to those that Microsoft builds into Datacenter. Unisys's Integrated Management System (IMS) is a partition-management application. The application's GUI displays the processes involved in assigning hardware components to partitions. The GUI uses an icon with two arrows forming a circle to indicate that a partition is running; an icon with one arrow forming a half-circle indicates that a partition is alive but not running.

Unisys also preinstalls a variety of scripts, some of which the smallest window in Figure 1 lists. You can run these scripts manually or combine scripts with monitoring functions to perform hardware-to-process affinity tasks. For example, you can combine a Performance Monitor alert with a script that will reassign processors to a process's I/O tasks when that process uses more I/O than the currently assigned processors can handle easily. Scripts can also run monitors to collect information about the environment. You can launch them manually at any time or run them in the background at startup. You can configure scripts to send alerts that suggest appropriate environment changes to administrators, or you can configure the scripts to make the changes automatically. Other Datacenter OEMs provide similar administrative tools. For more information about Datacenter OEMs, see Buyer's Guide, "Datacenter OEMs," page 79.

Another helpful feature is the snapshot you can take of the Datacenter system, at intervals you choose. Snapshots contain information about running programs, hardware utilization, and other aspects of the system's environment. Reviewing this information will help you decide how to optimize the system.

Playing with Datacenter
Like most of you, I don't own a Datacenter server, so I took the opportunity to walk through some tasks that Datacenter administrators might perform. I reassigned processors to processes manually and by using scripts. I also moved sub-pods from one partition to another.

Of course, I used a best-practices mentality and viewed the current states of partitions and groups before I made any changes. You can easily check system overviews and specific information about partitions and process groups. Figure 1 shows reports of performance and the processes and groups involved in moving processor sub-pods.

Reassigning processors to processes within a partition is an on-the-fly adjustment. In future Datacenter versions, Microsoft plans to overcome the OS's inability to perform on-the-fly reassignments of CPUs to partitions. Currently, reassigning a sub-pod to another partition causes the affected partitions to reboot; reassigning MSUs requires a complete system reboot.

By the end of the day, Datacenter's Performance Monitor environment reports impressed me more than the ease with which the server implements tasks. Incredulously, I observed the performance levels and the hardware you can adjust to maintain those levels. Heavy-duty database I/O and CPU utilization tests (from Microsoft SQL Server and Microsoft Exchange Server running individually and both at once) produced lower utilization numbers than I thought possible. Tasks that would severely stress my Win2K servers, and other servers more robust than mine, never produced utilization percentages beyond the teens on the ES7000 the lab technicians tested during my visit.

At the Unisys lab, I realized what scalability can mean. When I said this out loud, the technicians in the room were nice enough to avoid patronizing smiles and replies; all I heard was a subtle "Uh-huh."