email article printer friendly

Grid Computing Comes to Windows
PolyServe creates shared-everything clusters with Windows servers.
by Danielle Ruest and Nelson Ruest

August 24, 2004

Editor's Note: The authors did not have access to storage area network technology for this article. Therefore, this article is based on a demonstration of the product and is not a review or evaluation.

Everyone has heard of grid computing-the sharing of CPUs located in distributed computers to perform a set of operations-but no one had heard about it on Windows until PolyServe made a groundbreaking announcement at Microsoft's TechEd Conference in San Diego on May 24, 2004. That's the day PolyServe unveiled PolyServe Matrix Server for Windows (PMXS). PMXS creates shared-everything clusters on Windows, or as Polyserve calls it, "shared data" clusters. Previously these types of clusters were only available in Unix environments.

"This opens up a breadth of possibilities to Windows users," said Steve Norall, PolyServe's marketing director. "Now Windows users can use our technology to create clusters of up to 16 nodes from basically any hardware that is designed to run either Windows 2000 or Windows Server 2003."

PolyServe treats all node members as individual servers; therefore, it doesn't care if one node has four 2-GHz Intel CPUs while the other sports twin 900-MHz AMD CPUs. Because it uses a single binary for its operation, it also doesn't care if one node runs Windows 2000 and the other runs Windows Server 2003.

Another differentiator is that it uses a shared-everything model. This means each node has access to the same disk resources at the same time (see Figure 1). The advantage is if a service fails on one node, it can be immediately available on the other because the failover node does not need to take time to connect to the disk resource. This is one drawback of the Microsoft Cluster Service. Because it uses a shared-nothing model, each node must take the time to connect to the shared resource before picking up the failover from another node. The precious minutes this takes can mean phone calls to the help desk from disgruntled users. Not with PMXS.

Though PMXS works with any server hardware, it does have some hardware requirements. First, it needs access to a storage area network (SAN). Most SANs that work with Windows will work with PMXS. Two key hardware components are required, the host bus adapter (HBA) that connects a node to the SAN and the switch managing data throughput to the SAN.

"It's really nothing special because all SANs have these components," Norall said. In addition, PolyServe recommends two network interface cards (NIC) for each node: one for virtual services and client communications and one for private cluster communications. This is nothing new in a cluster configuration, though PMXS will work with only one. If two NICs are used, each one can serve as the failover for the other.

Two core components make PMXS tick. The first is a disk queue manager called the Distributed Lock Manager (DLM). This manages disk access for all nodes from a special membership partition stored on the SAN. To develop the DLM, PolyServe entered into a special partnership with Microsoft two years ago. Microsoft licensed special file service application programming interfaces to PolyServe to enable the company to delve deeply into the way Windows manages file access and create a working disk queue manager.