US 5,513,371 AGranted

Hierarchical interconnection network architecture for parallel processing, having interconnections between bit-addressible nodes based on address bit permutations

AssigneeInternational Business Machines Corporation

InventorsRobert E. Cypher, Jorge L. C. Sanz

Life of this patent

Estimating the term…

Loading patent…

Abstract

Two new classes of interconnection networks are described. The new classes of interconnection networks are referred to herein as the hierarchical shuffle-exchange (HSE) and hierarchical de Bruijn (HdB) networks. The new HSE and HdB networks are highly regular and scalable and are thus well suited to VSLI implementation. In addition, they can be adjusted to match any set of packaging constraints. These new networks are also efficient in supporting the execution of a wide range of algorithms on computers whose processors are interconnected via one of the networks fabricated in accordance with the teachings of the invention. Such computers, also contemplated by the invention, are referred to herein as HSE and HbB computers. Furthermore, methods for implementing the aforementioned wide range of algorithms, particularly those in the classes of Ascend and Descend algorithms, on the novel HSE and HdB computers, constitute a further aspect of the invention.

Claims10

Dependent claims

For implementing an Ascend algorithm having 2.sup.n data items, each of which has a unique integer ID having a value between 0 and 2.sup.n -1 inclusive, where n is an integer, wherein said Ascend algorithm has n stages, 0 through n-1, and further wherein at each state i, 0.ltoreq.i.ltoreq.n-1, processing is performed on each pair of data items with IDs whose binary representation differ only in bit position i, on an Hierarchical Shuffle-Exchange (HSE) computer having 2.sup.n processors, each of which has a unique ID having a value between 0 and 2.sup.n -1, inclusive wherein said HSE computer includes 2.sup.n-m packaging modules, where m is an integer and n>m, said method comprising the steps of:

(a) storing each data item j, where j is an integer having a value between 0 and 2.sup.n -1 inclusive, in processor j;
(b) performing the first m stages of said Ascend algorithm by utilizing exchange connections and local unshuffle connections of said HSE computer, the exchange connections being connections between each given one of the processors and a respective other one of the processors having an ID identical to that of the given one of the processors, except for a least significant bit opposite to that of the given one of the processors, the local unshuffle connections being connections between each given one of the processors and a respective other one of the processors having an ID identical to that of the given one of the processors, except that the m lowest significance bits of the ID of the other one of the processors, are those of the given one of the processors, subjected to a right rotate;
(c) repositioning the data items by utilizing global unshuffle connections of said HSE computer, the global unshuffle connections being connections between each given one of the processors and a respective other one of the processors having an ID identical to that of the given one of the processors, except that the bits of the ID of the other one of the processors are those of the given one of the processors, wherein the n-m most significant bits are in the n-m least significant bit positions, and the m least significant bits are in the m most significant bit positions;
and
(d) repeating steps (b) and (c) until all the n stages of the Ascend algorithm have been completed.

2 more elements

A method for implementing an Ascend algorithm having 2.sup.n+1 data items, each of which has a unique integer ID in the range of 0 through 2.sup.n+1 -1, where n is an integer, wherein said Ascend algorithm has n+1 stages, 0 through n, including first m+1 stages and further wherein at each stage i,0.ltoreq.i.ltoreq.n, processing is performed on each pair of data itmes with IDs whose binary representation differ only in bit position i, on an Hierarchical de Bruijn (HdB) computer having 2.sup.n processors, each of which has a unique ID having a value between 0 and 2.sup.n -1 inclusive, wherein HdB computer includes 2.sup.n-m packaging modules, where m is an integer and n>m, said method comprising the steps of:

(a) storing each data item j, where j is an integer having a value between 0 and 2.sup.n+1 -1 inclusive, in the processor whose ID is a largest integer less than or equal to j/2;
(b) performing the first m+1 stages of said Ascend algorithm by utilizing local unshuffle-replace connections of said HdB computer, the local unshuffle-replace connections being connections between each given one of the processors and a respective other one of the processors having an ID identical to that of the given one of the processors, except that the m lowest significance bits of the ID of the other one of the processors, including the i-th bit, are those of the given one of the processors, subjected to a right shift, the least significant bit being discarded, and a replace bit being inserted at the (m-1)th bit position;
(c) repositioning the data items by utilizing global unshuffle-replace connections of said HdB computer, the global unshuffle-replace connections being connections between each given one of the processors and a respective other one of the processors having an ID identical to that of the given one of the processors, except that the bits of the ID of the other one of the processors are those of the given one of the processors, wherein the n-m-1 most significant bits are in the n-m-1 least significant bit positions, the m least significant bits are in the m most significant bit positions, and a replace bit is in the (m+1)th most significant bit position;
and
(d) repeating steps (b) and (c) until all the n+1 stages of the Ascend algorithm have been completed.

2 more elements

A method for implementing a Descent algorithm having 2.sup.n data items, each of which has a unique integer ID having a value between 0 and 2.sup.n -1 inclusive;

where n is an integer, wherein said Descent algorithm has n stages, 0 through n-1, and further wherein at each stage i, 0.ltoreq.i.ltoreq.n-1, processing is performed on each pair of data items with IDs whose binary representation differ only in bit position i, on an Hierarchical Shuffle-Exchange (HSE) computer having 2.sup.n processors, each of which has a unique ID having a value between 0 and 2.sup.n -1 inclusive wherein said HSE computer includes 2.sup.n-m packaging modules, where m is an integer and n>m, said method comprising the steps of:
(a) storing each data item j, where j is an integer having a value between 0 and 2.sup.n -1, inclusive, in processor j;
(b) repositioning the data items by utilizing global shuffle connections of said HSE computer, the global shuffle connections being connections between each given one of the processors and a respective other one of the processors having an ID identical to that of the given one of the processors, except that the bits of the ID of the other one of the processors are those of the given one of the processors, wherein the m most significant bits are in the m least significant bit positions, and the n-m least significant bits are in the n-m most significant bit positions;
(c) performing the stages n-1 through n-m of said Descend algorithm by utilizing local shuffle connectors and exchange connections of said HSE computer, the local shuffle connections being connections between each given one of the processors and a respective other one of the processors having an ID identical to that of the given one of the processors, except that the m lowest significance bits of the ID of the other one of the processors, are those of the given one of the processors, subjected to a left rotate, the exchange connections being connections between each given one of the processors and a respective other one of the processors having an ID identical to that of the given one of the processors, except for a least significant bit opposite to that of the given one of the processors;
and
(d) repeating steps of (b) and (c) until all the n stages of the Descend algorithm have been completed.

3 more elements

A method for implementing a Descend algorithm having 2.sup.n+1 data items, each of which has a unique integer ID having a value between 0 and 2.sup.n+1 inclusive, where n is an integer, wherein said Descend algorithm has n+1 stages, 0 through n, and further wherein at each stage i, 0.ltoreq.i.ltoreq.n, processing is performed on each pair of data items with IDs whose binary representation differ only in bit position i, on an Hierarchical de Bruijn (HdB) computer having 2.sup.n processors, each of which has a unique ID having a value between 0 and 2.sup.n -1 inclusive, wherein said HdB computer includes 2.sup.n-m packaging modules, where m is an integer and n>m, said method comprising the steps of:

(a) storing each data item j, where j is an integer having a value between 0 and 2.sup.n+1 -1, inclusive, in the processor whose ID is a largest integer less than or equal to j/2;
(b) repositioning the data items by utilizing global shuffle-replace connections of said HdB computer, the global shuffle-replace connections being connections between each given one of the processors and a respective other one of the processors having an ID identical to that of the given one of the processors, except that the bits of the ID of the other one of the processors are those of the given one of the processors, wherein the n-m-1 least significant bits are in the n-m-1 most significant bit positions, the m most significant bits are in the m least significant bit positions, and a replace bit is in the (m+1)th least significant bit position;
(c) performing stages n through n-m of said Descend algorithm by utilizing local shuffle-replace connections of said HdB computer, the local shuffle-replace connections being connections between each given one of the processors and a respective other one of the processors having an ID identical to that of the given one of the processors, except that the m lowest significance bits of the ID of the other one of the processors, including the i-th bit, are those of the given one of the processors, subjected to a left shift, the most significant of the m lowest significant bits being discarded, and a replace bit being inserted at the least significant bit position;
and
(d) repeating steps (b) and (c) until all the n+1 stages of the Descend algorithm have been completed.

2 more elements

Dependencies are inferred from the claim text.