A VMScluster is a computer cluster involving a group of computers running the OpenVMS operating system. Whereas tightly coupled multiprocessor systems run a single copy of the operating system, a VMScluster is loosely coupled: each machine runs its own copy of OpenVMS, but the disk storage, lock manager, and security domain are all cluster-wide. Machines can join or leave a VMScluster without affecting the rest of the cluster. For enhanced availability, VMSclusters support the use of dual-ported disks connected to two machines or storage controllers simultaneously. With OpenVMS now ported to Alpha and IA-64 machines, the facility originally named VAXclustering was renamed to VMSclustering.
Digital Equipment Corporation first announced VAXclusters in May 1983. At this stage, clustering required specialised communications hardware, as well as some major changes to low-level subsystems in VMS. The software and hardware were designed jointly.
At the center of each cluster was a star coupler, to which every node (computer) and data storage device in the cluster was connected by one or two pairs of CI cables. ("CI" stands for Computer Interconnect.) Each pair of cables had a transmission rate of 70 megabits per second, a high speed for that era. Using two pairs gave an aggregate transmission rate of 140 megabits per second, with redundancy in case one cable failed; the star couplers also had redundant wiring for better availability.
Each CI cable connected to its computer via a CI Port, which could send and receive packets without any CPU involvement. To send a packet, a CPU had only to create a small data structure in memory and append it to a "send" queue; similarly, the CI Port would append each incoming message to a "receive" queue. Tests showed that a VAX-11/780 could send and receive 3000 messages per second, even though it was nominally a 1-MIPS machine. The closely related Mass Storage Control Protocol (MSCP) allowed similarly high performance from the mass storage subsystem. In addition, MSCP packets were very easily transported over the CI allowing remote access to storage devices.
VAXclustering was the first clustering system to achieve commercial success, and was a major selling point for VAX systems.
In 1986, DEC added VAXclustering support to their MicroVAX minicomputers, running over Ethernet instead of special-purpose hardware. While not giving the high-availability advantages of the CI hardware, these Local Area VAXclusters provided an attractive expansion path for buyers of low-end minicomputers.
Later versions of OpenVMS (V5.0 and later) supported "mixed interconnect" VAXclusters (using both CI and Ethernet), and VAXclustering over DSSI (Digital Systems and Storage Interconnect) and FDDI, among other transports. Eventually, as high-bandwidth wide area networking became available, clustering was extended to allow satellite data links and long-distance terrestrial links. This allowed the creation of disaster-tolerant clusters; by locating the single VAXcluster in several diverse geographical areas, the cluster could survive infrastructure failures and natural disasters.
VAXclustering was greatly aided by the introduction of terminal servers using the LAT protocol. By allowing ordinary serial terminals to access the host nodes via Ethernet, it became possible for any terminal to rapidly and easily connect to any host node. This made it much simpler to accomplish fail over of the user terminals from one node of the cluster to another.
Eventually, VAXclusters reached the point where the cluster as a whole essentially never went down. Rolling upgrades even allowed the system operators to upgrade the OpenVMS system software, shutting down, upgrading, and rebooting individual nodes while the cluster as a whole continued processing. Cluster uptimes are frequently measured in years with the current longest uptime being at least sixteen years.
As mentioned above, OpenVMS now also runs on Alpha and IA-64 systems, so the term VAXcluster has been replaced by VMScluster. With Gigabit Ethernet now common and 10-gigabit Ethernet being introduced, standard networking cables and cards are quite sufficient to support VMSclustering.
- Nancy P. Kronenberg, Henry M. Levy, William D. Strecker, "VAXcluster: a closely-coupled distributed system". ACM Transactions on Computer Systems 4 (2), 1986
- DIGITAL Technical Journal 3 (3), 1991 This issue was devoted to VAXclusters and FDDI networking. (Archived as PDF files.)