372 Part 2½ Regions of Computer Space Section 4 ½Multiple-Processor Systems
can quickly reinitialize itself, omitting troublesome components. This
approach is especially suitable for applications in which brief outages
can be tolerated and where overall correctness can be ensured by other
II. Pluribus System Architecture
The Pluribus may be characterized as a symmetric, tightly coupled multiprocessor, designed to be flexible and highly modular. Modules are physically isolated to protect against common failures, and a form of distributed switch is employed for intermodule communications. In this section, we discuss these characteristics and describe the hardware architecture of the Pluribus.
A. Major Design Decisions
In order to make the basic operation of the Pluribus clearer, it is useful to examine some of the major design decisions that have directed its development, and to consider those decisions in the context of other options for multiprocessor system design. We have identified three areas which we believe are key aspects of the Pluribus approach to multiprocessing, each of which is considered in greater detail below.
Processor Symmetry One dimension of multiprocessing involves the degree of inter-processor symmetry within the system [Enslow, 1974, p. 83]. In this dimension, one extreme might be a typical general purpose computer system, including a central processor, a front-end processor, and perhaps one or more channel processors. Such an asymmetric system is relatively inflexible in power since increasing its central processing capacity requires the introduction of a more powerful central processor. Building redundancy into an asymmetric system can be expensive, since replication of all critical resources involves duplicating virtually the whole machine.
At the other extreme are systems like the Pluribus in which all processors are identical. In such systems, the advantages of redundancy and flexibility are much easier to achieve since they include only one type of processing unit. Even without explicit redundancy, a symmetric system can provide graceful degradation of throughput when a processing element fails. Pluribus systems which are sized for fully redundant operation include just one extra processing module; thus the degradation which results from failure of any processing module consists only of a loss of excess throughput capacity.
Processor Coupling Another multiprocessing dimension is the level at which processors cooperate to accomplish overall system requirements. At one extreme the processors might run totally separate programs under the direction of a supervisor program, communicating only at arm's length. Such processors may be described as "loosely coupled" [Enslow, 1974, p. 15]. At the other extreme, which is characterized by array processors such as ILLIAC IV [Barnes et al., 19681, the processors run in lockstep, with a single program operating simultaneously on a number of data streams. The Pluribus lies between these extremes. Its processors are tightly coupled in the sense that all processors can access all system resources and perform all parts of the operational program; they operate independently except for necessary software interlocks on specific I/O devices and data structures.
Flexibility Although one of the goals in the creation of the Pluribus was to develop a machine with high throughput, this goal was complemented by the need for a smaller, cheaper machine with relatively low throughput. Similarly, although the Pluribus was conceived as having at least two of every resource to permit recovery after failures, it was also clear that not all applications required or could afford a fully redundant system. Thus it was desirable for the architecture to be flexible in at least two ways:
The size-flexibility goal was to smooth large incremental steps in the
cost-performance curve by utilizing a highly modular design, which could
provide processing capacity well beyond our anticipated needs. Flexibility
in the area of fault-tolerance and fault-recovery was a related goal, since
the need for fault-tolerance involves primarily economic considerations
and we wanted to allow our customers to select fault-tolerance features
independent of their throughput requirements. Also implied in each of these
goals was the requirement for easy expansion to meet changing requirements.
previous | contents | next