Algorithms and Parallel Computing

3.3 SHARED-MEMORY MULTIPROCESSORS (UNIFORM MEMORY ACCESS [UMA])

Shared-memory processors are popular due to their simple and general programming model, which allows simple development of parallel software that supports sharing of code and data [27]. Another name for shared memory processors is parallel random access machine (PRAM). The shared-memory or shared-address space is used as a means for communication between the processors. All the processors in the shared memory architecture can access the same address space of a common memory through an interconnection network as shown in Fig. 3.1a. Typically, that interconnection network is a bus, but for larger systems, a network replaces the bus to improve performance. The performance we are referring to is the amount of processor/memory accesses that can be performed per unit time (throughput) and the time delay between a processor requesting memory access and the time when that request is granted (delay). Examples of the types of interconnection networks and their performance analysis can be found in Reference 28.

Figure 3.1 Shared-memory multiprocessor architecture (PRAM). (a) The processors are connected to the shared memory using a single bus. (b) The processors and memory modules are connected using an interconnection network.

We can immediately see that the memory bandwidth becomes the system bottleneck since only one processor ...