Cluster of workstations for solving compute-intensive applications by exchanging interim computation results using a two phase communication protocol

• US 5,944,779 A
• Filed: 07/02/1996
• Issued: 08/31/1999
• Est. Priority Date: 07/02/1996
• Status: Expired due to Fees

• Current Assignee: Compbionics Incorporated
• Original Assignee: Compbionics Incorporated
• Inventors: Blum, Edward K.
• Primary Examiner(s): Maung, Zarni
• Assistant Examiner(s): CALDWELL, ANDREW T
• Application Number:
  US08/675,394
• Time in Patent Office: 1,155 Days
• Field of Search: 395/672, 395/200.62, 395/800.28, 395/800.29, 395/800.3, 395/800.31, 395/200.79, 395/306, 395/307, 395/308, 395/309, 395/310, 395/311, 395/312 , 395/800.16, 395/800.17, 395/800.18, 395/800.19, 395/800.2, 395/800.21, 395/800.22, 395/200.31
• US Class Current: 709/201
• CPC Class Codes: G06F 9/466 Transaction processing

Abstract

A system and method for connecting computer workstations in clusters to perform parallel-distributed processing with respect to compute-intensive applications are provided. Commodity computers/workstations and commodity network hardware are arranged to form unique-architecture building blocks (clusters) which may then act as supernodes in larger scale cluster systems. An integrated ROCC+ (Reduced Overhead Cluster Communication) message passing software system provides unique communication logic for efficient implementation of collective message passing operations between each node and supernode. According to a preferred embodiment, each building block comprises two ethernet segments and four nodes (e.g., workstations), two of which are connected by Network Interface Cards (NICs) to both segments with the remaining nodes each connected to a respective one of the segments. The ROCC+ software establishes communication links within and between the basic building blocks so as to eliminate collisions (contention) on segments while allowing for parallel (concurrent) message passing between nodes and supernodes.

Claims

1. A computer network architecture system comprising:

a first segment;

a second segment;

a first computer having two network interface cards associated therewith;

said first computer being connected to said first network segment through a first of said two network interface cards, and said first computer being connected to said second segment through a second of said two network interface cards;

a second computer having two network interface cards associated therewith;

said second computer being connected to said first network segment through a first of said two network interface cards, and said second computer being connected to said second segment through a second of said two network interface cards;

a third computer having one network interface card associated therewith, said third computer only being connected to said first segment through said one network interface cards; and

a fourth computer having one network interface card associated therewith, said fourth computer only being connected to said second segment through said one network interface cards;

wherein said first, second, third, and fourth computers cooperate to solve an application problem by each calculating interim results and then exchanging messages including the interim results using a two phase message passing protocol in which;

during a first phase of said two-phase message passing protocol, said first computer and said third computer exchange messages on said first network segment and said second and said fourth computer exchange messages on said second network segment;

during a second phase of said two-phase message passing protocol, said first computer and said fourth computer exchange messages on said second network segment and said second computer and said third computer exchange messages on said first network segment.

2. The system of claim 1 wherein said first, second, third and fourth computers, respectively, comprise personal computers.

3. The system of claim 1 wherein said first, second, third and fourth computers, respectively, comprise workstations.

4. The system of claim 1 wherein said first and second segments, respectively, comprise ethernet cable.

5. The system of claim 1 wherein said first and second segments, respectively, comprise hubs.

6. A computer network architecture system comprising m supernodes, wherein m is an integer greater than one, each of said m supernodes further comprising:

a first segment;

a second segment;

a first computer having at least two network interface cards associated therewith, said first computer being connected to said first segment through a first of said at least two network interface cards, and said first computer being connected to said second segment through a second of said at least two network interface cards;

a second computer having at least two network interface cards associated therewith, said second computer being connected to said first segment through a first of said at least two network interface cards, and said second computer being connected to said second segment through a second of said at least two network interface cards;

a third computer having at least one network interface card associated therewith, said third computer being connected to said first segment through a first of said at least one network interface card; and

,

a fourth computer having at least one network interface card associated therewith, said fourth computer being connected to said second segment through a first of said at least one network interface card;

means for interconnecting said m supernodes; and

wherein said means for interconnecting comprise a third network interface card associated with each of said first and second computers of each of said m supernodes for use in connecting said m supernodes in a daisy-chain.

7. The system of claim 6 wherein said third network interface card of said first computer of the ith supernode, wherein i=1 to m, is connected to said first segment of the i+1th supernode, said third network interface card of said second computer of said ith supernode is connected to said second segment of said i+1th supernode, and the first supernode is the i+1th supernode for i=m.

8. The system of claim 7 further comprising means for establishing communication links and passing messages between said first, second, third and fourth computers of each of said m supernodes, respectively, and further between each of said m supernodes so as to effect parallel-distributed processing.

9. The system of claim 6 wherein said third network interface card of said first computer of the ith supernode, wherein i=1 to m, is connected to said first segment of the i+1th supernode, said third network interface card of said second computer of said ith supernode is connected to said second segment of the i-1th supernode, and the first supernode is the i+1th supernode for i=m, and the mth supernode is the i-1th supernode for i=1.

10. The system of claim 6 further comprising means for establishing communication links and passing messages between said first, second, third and fourth computers of each of said m supernodes, respectively, and further between each of said m supernodes so as to effect parallel-distributed processing.

11. The system of claim 6 wherein said means for interconnecting comprise a second network interface card associated with each of said third and fourth computers of said m supernodes for use in connecting said m supernodes in a daisy-chain.

12. The system of claim 11 wherein said second network interface card of said third computer of the ith supernode, wherein i=1 to m, is connected to said first segment of the i+1th supernode, said second network interface card of said fourth computer of said ith supernode is connected to said second segment of said i+1th supernode, and the first supernode is the i+1th supernode for i=m.

13. The system of claim 12 further comprising means for establishing communication links and passing messages between said first, second, third and fourth computers of each of said m supernodes, respectively, and further between each of said m supernodes so as to effect parallel-distributed processing.

14. The system of claim 11 wherein said second network interface card of said third computer of the ith supernode, wherein i=1 to m, is connected to said first segment of the i+1th supernode, said second network interface card of said fourth computer of said ith supernode is connected to said second segment of the i-1th supernode, and the first supernode is the i+1th supernode for i=m, and the mth supernode is the i-1th supernode for i=1.

15. The system of claim 14 further comprising means for establishing communication links and passing messages between said first, second, third and fourth computers of each of said m supernodes, respectively, and further between each of said m supernodes so as to effect parallel-distributed processing.

16. The system of claim 6 wherein said means for interconnecting comprise a third and fourth network interface card associated with each of said first and second computers of each of said m supernodes for use in connecting said m supernodes in a torus architecture.

17. The system of claim 6 wherein said means for interconnecting comprise a third network interface card associated with each of said first and second computers of each of said m supernodes, and a second network interface card associated with each of said third and fourth computers of each of said m supernodes, for use in connecting said m supernodes in a torus architecture.

18. In a computer network architecture system comprising at least one supernode comprising first and second segments, first and second nodes each connected to said first and second segments, a third node connected to said first segment, and a fourth node connected to said second segment, a method for parallel-distributed processing comprising the steps of:

simultaneously processing data in each of said first, second, third and fourth nodes, respectively;

in a first phase, passing data on said first segment from said first node to said third node and then from said third node to said first node while simultaneously passing data on said second segment from said second node to said fourth node and then from said fourth node to said second node; and

,

in a second phase;

passing data on said second segment from said first node to said fourth node and then from said fourth node to said first node while simultaneously passing data on said first segment from said second node to said third node and then from said third node to said second node.

19. The method of claim 18 further comprising the step of:

systematically passing data between pairs of nodes wherein each node of a respective pair of nodes is associated with a different supernode.

20. The method of claim 19 wherein the step of systematically passing data further comprises the steps of:

passing data between successive supernodes in parallel and on distinct segments in a shift phase; and

,

passing data between nodes within each supernode in parallel and on distinct segments in a merge phase.

21. The method of claim 20 further comprising the step of:

alternately repeating said shift phase and said merge phase until all data is passed between all nodes and all supernodes.

22. In a computer network architecture system comprising at least one supermode consisting essentially of:

first and second segments, first and second nodes each connected to said first and second segments, a third node connected to said first segment, and a fourth node connected to said second segment, a method for parallel-distributed processing comprising the steps of;

processing data within each of said first, second, third and fourth nodes in parallel; and

,

systematically passing messages comprising intermediate results of calculations between pairs of nodes in such manner as to provide all messages to all nodes.

23. A system for parallel-distributed processing made up of plurality of supernodes, each consisting essentially of:

first and second segments;

first and second nodes each connected to said first and second segments,

a third node connected to said first segment;

a fourth node connected to said second segment; and

,

means for establishing communication links and passing messages between each of said first, second, third and fourth nodes, said means providing for concurrent message passing between pairs of said nodes while avoiding contention on said first and second segments during message passing, and said means resulting in communication of processing results of each of said first, second, third and fourth nodes with the other three of said nodes, respectively.

24. The system of claim 23 wherein said means for establishing communication links and passing messages comprises interconnected hardware components and associated software.

25. A computer network architecture system made up of a plurality of computers arranged in supernodes, wherein each supernode consists essentially of:

a first segment;

a second segment;

a first computer having two network interface cards associated therewith, said first computer being connected to said first segment through a first of said two network interface cards, and said first computer being connected to said second segment through a second of said two network interface cards;

a second computer having two network interface cards associated therewith, said second computer being connected to said first segment through a first of said two network interface cards, and said second computer being connected to said second segment through a second of said network interface cards;

a third computer having one network interface card associated therewith, said third computer only being connected to said first segment through a first of said network interface cards; and

,

a fourth computer having one network interface card associated therewith, said fourth computer only being connected to said second segment through a first of said one network interface cards.