What Wolfpack Means for Parallel Computing

Microsoft's clustering enhancement for Windows NT Server, code-named Wolfpack, aims to extend the operating system's symmetric multiprocessing (SMP) capabilities in two key areas: high availability (i.e., fault tolerance) and scalability. Mark Wood, product manager for Wolfpack, is quick to point out that "Wolfpack clusters aren't a substitute for RAID or regular backups -- they're a part of server protection."

Wolfpack will be introduced in two phases. The first phase, which should be available about the time you read this, emphasizes high availability for traditional server applications. It supports only twin-processor clusters, with a fail-over capability so that if one processor fails, the system will recover automatically by migrating all its workload onto the other CPU. These clusters are initially going to be built by connecting two systems through a SCSI bus using SCSI-2's multi-initiator connection. Because not all SCSI host adapters support the SCSI Release/Reserve command the same way, Microsoft has created a hardware certification program. Initially, Microsoft will be certifying only complete configurations (including servers, storage, and the interconnect). "When it becomes practical," adds Wood, "we will certify individual components." But at first, Microsoft wants to make sure things go smoothly.

Once you've set up your hardware, you use a wizard in Wolfpack Manager to create sets of applications that you want to cluster. These sets are called cluster groups. Each cluster group gets an IP address. Using IP addresses like this enables off-the-shelf applications to take advantage of Wolfpack without any special client software.

Phase two of Wolfpack will be more interesting from a parallel computing viewpoint. The technology will support clusters of more than two nodes over interconnects such as Fibre Channel. In addition, Wolfpack will offer dynamic load balancing over and above fail-over. This will enable the OS to migrate work from one CPU to another to maximize throughput rather than to just recover from a crash.

Beyond this stage, NT with Wolfpack will become a tempting choice of OS around which to design a clustered-SMP computer since it will support all the important high-speed interconnect technologies like asynchronous transfer mode (ATM) and Fibre Channel thro ugh its standard driver model. Phase two, scheduled for 1998, will support SQL Server in shared-nothing disk configurations but will not contain a distributed lock manager (DLM) for shared disk access (which Microsoft was considering at one time). However, many parallelized enterprise databases like Oracle have already implemented distributed locking at the application level.