previous | contents | next

Chapter 27 ÷ The ILLIAC IV computer 321

processing elements in the array. This eliminates the cost and complexity for decoding and timing circuits in each element.

In addition, an index register and address adder are provided with each processing element, so that the final operand address ai for element i is determined as follows:

ai = a + (b) + (ci)

where a is the base address specified in the instruction, (b) is the contents of a central index register in the control unit, and (ci) is the contents of the local index register of the processing element i. This independence in operand addressing is very effective for handling rows and columns of matrices and other multidimensional data structures [Kuck, 1968].

Mode control and data conditional operations

Although the goal of the ILLIAC IV structure is to be able to control the processing of a number of data streams with a single instruction stream, it is sometimes necessary to exclude some data streams or to process them differently. This is accomplished by providing each processor with an ENABLE flip-flop whose value controls the instruction execution at the processor level.

The ENABLE bit is part of a test result register in each processor which holds the results of tests conditional on local data. Thus in ILLIAC IV the data conditional jumps of conventional computers are accomplished by processor tests which enable or disable local execution of subsequent commands in the instruction stream.

Routing

Each processing element i in the ILLIAC IV has data routing connections to 4 of its neighbors, processors i + 1, i - 1, i + 8, and i - 8. End connection is end around so that, for a single array, processor 63 connects to processors 0, 62, 7, and 55.

Interprocessor data transmissions of arbitrary distance are accomplished by a sequence of routings within a single instruction. For a 64-processor array the maximum number of routing steps required is 7; the average overall possible distances is 4. In actual programs, routing by distance 1 is most common and distances greater than 2 are rare.

Common operand broadcasting

Constants or other operands used in common by all the processors are fetched and stored locally by the central control and broadcast to the processors in conjunction with the instruction using them. This has several advantages: (1) it reduces the memory used for storage of program constants, and (2) it permits overlap of common operand fetches with other operations.

Processor partitioning

Many computations do not require the full 64-bit precision of the processors. To make more efficient use of the hardware and speed up computations, each processor may be partitioned into either two 32-bit or eight 8-bit subprocessors, to yield 512 32-bit or 2048 8-bit subprocessors for the entire ILLIAC IV set.

The subprocessors are not completely independent in that they share a common index register and the 64-bit data routing paths. The 32-bit subprocessors have separate enabled/disabled modes for indexing and data routing; the 8-bit subprocessors do not.

Array partitioning

The 256 elements of ILLIAC IV are grouped into four separate subarrays of 64 processors, each subarray having its own control unit and capable of independent processing. The subarrays may be dynamically united to form two arrays of 128 processors or one array of 256 processors. The following advantages are obtained.

1 Programs with moderately dimensioned vector or matrix variables can be more efficiently matched to the array size.

2 Failure of any subarray does not preclude continued processing by the others.

This paper summarizes the structure of the entire ILLIAC IV system. Programming techniques and data structures for ILLIAC IV are covered in a paper by Kuck [1968].

ILLIAC IV structure

The organization of the ILLIAC IV system is indicated in Fig. 1. The individual processing elements (PEs) are grouped in four arrays, each containing 64 elements and a control unit (CU). The four arrays may be connected together under program control to permit multiprocessing or single-processing operation. The system program resides in a general-purpose computer, a Burroughs B 6500, which supervises program loading, array configuration changes, and I/O operations internal to the ILLIAC IV system and to the external world. To provide backup memory for the ILLIAC IV arrays, a large parallel-access disk system (10 bits, 109 bit per second access rate, 40-ms maximum latency) is directly coupled to the arrays. There is also provision for real-time data connections directly to the ILLIAC IV arrays.

previous | contents | next