суперкомпьютер семейства ILLIAC (ILLInois Automated Computer), страница 5

However, it is expected that such formatting or preprocessing canbe done more expeditiously on the ILLlAC IV array, which will have theadditional advantage of avoiding the potentially inefficient situation ofhaving the array idle waiting on the B 6700.Array processing starts by taking all the data collected on the disk andmoving only that portion required by the beginning of the program into thearray memory; that is, the PEMs are collectively treated as a single memorybank (path "2" in the figure). This move is under the control of the MCPin the B 6700.To make all data transfers efficient, an Optimizer is included in the I10Subsystem so that many independent blocks of data, which could betransferred together, can be included into a single transfer operation with asingle access time, without the necessity of being combined under a single-As processing proceeds in the array, various data objects will be created inthe array, and sent to disk (path "3" of the figure).

These include restartinformation, overlayed information in the case of problems that do not fitin main memory, and intermediate results. In the present design of ILLIACIV, the B 6700 MCP has the responsibility of disk space assignment, andsuch actions are initiated by an interrupt to the B 6700 MCP which issuesthe control words to the 110 Subsystem.At the end of data manipulation, all results of array processing will be ondisk, and outputs destined for users are transferred, via BIOM, toperipherals via the B 6700 multiplexor. Outputs destined for use by jobsfor subsequent execution on the array are transferred to the bulkstore. These transfers are under the control of the B 6700 MCP.Use of the ILLIAC LV may, in some cases, involve postprocessing executedby programs in the B 6700.

In general, use of the ILLlAC IV array forpostprocessing is encouraged for the same reasons as for preprocessing.The flow of both data and instructions is essentially identical up to thepoint of being inserted into the PEMs. The one exception is inpreprocessing of data recommended for execution in the array rather thanin the B 6700. The corresponding action on program instructions iscompiling which will be done on the B 6700 for the program to be run onthe array.Once in the array processor, however, flow of data and instructions i sdifferent. Data generally is segregated into the individual ProcessingElement Memories, and is fetched to a Processing Element from i t s ownProcessing Element Memory, and restored into that Processing ElementMemory.

Two mechanisms break through this segregation of data intoseparate Processing ElementMemories. The first is called"routing", the capability of theProcessing Elements to shift dataa mo ng t hemselves. The secondmechanism lies in the Control Unitwhich sees all of the Processing Element Memories as one singlememory bank. The Control Unit (a65th, different processor, itself) canfetch any word from this memorybank.

In some cases it will "broadcast" such a word of data in parallelto all Processing Elements.Programs stored in the memorybank are seen by the ControlUnit. Instructions are fetched inblocks of 16 instructions at a time,and fed, one instruction at a time,i n t o the instruction decodingm a c h i n e r y o f the ControlUnit. Although 64 data streamscan be worked on simultaneously,- one going between each Processing Element and i t s ProcessingElement Memory, o n l y oneinstruction stream is being decoded(that instruction stream beingfetched to the Control Unit andexecuted from there).

lnstructionsare expected normally to be readonly, and to be stored in the first6.25% of memory addresses whereprotection against writing is available.PERIPHERALSB 6700 SYSTEM QCONTROLBULK STORAGE I-Data-Flow in ILLIAC I VSOFTWAREThe l L L l AC IV software currently includes a macro-assembler, aFORTRAN compiler, an operating system, and a basic library of intrinsicfunctions and utility routines. A brief description of each follows:ASSEMBLER - The I L L I A C I V assembly languagecorresponds directly to the ILLlAC IV set of instructions. Inaddition, there is a set of pseudo-opcodes which handles suchactivities as allotting storage and loading data in the ProcessingElement* Memories.

The assembler also has numerous controloptions for simplifying the updating files.FORTRAN COMPILER - ILLIAC FORTRAN is an extensionof ANSI FORTRAN. The additional syntax makes it possibleto operate explicitly on vectors and to express the kind ofparallelism exploitable on the lLLlAC IV.To distinguish operations or vectors from those on subscriptedvariables, the concept of a binary vector has beenintroduced.

A binary vector is a string of bits; the number ofbits being equal to the length of the vector with which it willbe used. The bits in the binary vector can be set and reseteither explicitly or implicitly by logical operations. Whenoperations on a vector ararequired, a binary vector is used as asubscript. Code is emitted by the compiler so that the statedoperation is performed on all those elements of the vectorwhose corresponding bits are set in the binary vector.The width of a vector is immaterial.

As far as the programmeris concerned, the operation is done simultaneously on allelements specified by the binary vector.IL L lAC FORTRAN affords highly sophisticated matrixmanipulations through its vector operations capability. Thecolumns t o be manipulated can be selected by placing a binaryvector in the column subscript position. A specific row can beselected by using a scalar in the row subscript position; adifferent row element for each of the columns can be selectedby placing an integer vector in the row subscript position. Thisinteger vector would contain a row number corresponding toeach column.

Thus, for example, it is possible to operate onevery other member of the diagonal of a matrix in oneFORTRAN statement.To aid t h e programmer further, the compiler allowssubprograms to be compiled separately and assembler code tobe interspersed with the ILLIAC FORTRAN code. A widerange of file maintenance features is also included.INTRINSICS AND UTl LlTlES - A library of the standardintrinsic functions has been written. Included are thetrigonometric functions, the natural logarithm, theexponential, and square root. They all have an accuracy of 15to 16 octal characters.

Many utility routines have been adaptedfor the ILLIAC IV. Among these are routines to handleNewton-Rapheson solutions, eigen-value problems, fast Fouriertransforms, and Monte Carlo techniques.OPERATING SYSTEM -The operating system controls thelLLlAC IV environment. The operating system prepares a job,schedules it, moves its data and its program in and out of thearray, and handles all ILLIAC disk transfers. The operatingsystem can control the tasks for several jobs simultaneously..