Message transfer system and method for parallel computer with message transfers being scheduled by skew and roll functions to avoid bottlenecks

• US 5,617,538 A
• Filed: 05/26/1994
• Issued: 04/01/1997
• Est. Priority Date: 07/02/1991
• Status: Expired due to Fees

• Current Assignee: TM Patents LP
• Original Assignee: TM Patents LP
• Inventors: Heller, Steven K.
• Primary Examiner(s): Lim, Krisna
• Application Number:
  US08/249,653
• Time in Patent Office: 1,041 Days
• Field of Search: 395/200, 395/325, 395/800, 395/200.02, 395/823, 370/60, 370/92, 370/93, 340/825.52
• US Class Current: 709/245
• CPC Class Codes: G06F 15/17368 non hierarchical topologies

Abstract

A computer system comprises a plurality of processing elements interconnected by a communications network. The communications network has a series of network addresses each identifying a location in the network, each processing element has an associated network address in the series. The communications network transfers messages transmitted by the processing elements in accordance with a respective address portion associated with each message, thereby to transfer the messages among the processing elements. Each processing element includes a message generator and a message transmitter. The message generator generates, during a message transfer operation, a series of messages for transmission over the communications network to others of the processing elements in the system, each message having an address portion whose contents enable the communications network to transfer the message from the processing element generating the message to a processing element to receive the message. The message generator generates the messages such that address portions of the successive messages in the series are transferred to processing elements having successive network addresses. The message transmitter iteratively transmits the series of messages generated by the message generator, the message transmitters of the processing elements which have successive network addresses selecting, as initial messages to be transmitted, messages having address portions to be transmitted to processing elements which have succeeding network addresses, thereby to effect a skewing of the messages transferred over the communications network.

Claims

1. A computer system comprising a plurality of processing elements interconnected by a communications network, the communications network having a series of network addresses each associated with one of said processing elements,

A. the communications network transferring messages transmitted by the processing elements in accordance with an address portion associated with each message, each address portion identifying a network address, the communications network transferring each message to one of said processing elements whose network address is identified by the message'"'"'s address portion;

B. said processing elements during a message transfer operation, generating a series of messages and transmitting them over the communications network, each of said processing elements;

i. transmitting a first message whose address portion contains a network address which is a mathematical function of a skew parameter value and the processing element'"'"'s network address, and

ii. after transmitting the first message, transmitting successive messages having address portions containing successive network addresses following the first message'"'"'s network address in the series,

such that successive messages are associated with respective address portions which enable the communications network to transfer the messages to processing elements which have successive network addresses.

2. A system as defined in claim 1 in which the mathematical function is S₀ =Up mod M where S₀ is related to the network address for the address portion associated with the first message, "U" corresponds to the skew parameter value, "p" corresponds to the network address of the processing element generating the message, "M" corresponds to the number of messages to be transmitted by the processing element during the message transfer operation, and "mod" refers to a modulo function.

3. A computer system defined in claim 1 in which the series of network addresses has a network address range from a lower network address value to an upper network address value, with the series of network addresses sequencing from the lower network address value toward the upper network address value, p1 A. after successively transmitting messages respectively associated with address portions identifying network addresses in the portion of the network address range from the processing element'"'"'s network address toward the upper network address value, each processing element continues successively transmitting messages respectively associated with address portions identifying network addresses in the portion of the network address range from the lower network address value toward the upper network address value.

4. A computer system as defined in claim 1 in which, for each processing element, the address portion associated with each message after said first message contains a network address which is a second mathematical function of a roll parameter value and a network address of the address portion associated with a message previously transmitted by said each processing element.

6. A system as defined in claim 1 in which each processing element includes:

A. a message generator for generating the series of messages to be transmitted during a message transfer operation, the message generator ordering the messages such that successive messages in the series are to be transferred to processing elements having successive network addresses; and

B. a message transmitter for successively transmitting the series of messages generated by the message generator, the message transmitter selecting, as an initial messages to be transmitted, a message having a message address which is the mathematical function of the skew parameter and the processing element'"'"'s network address.

7. A system as defined in claim 6 in which, in each processing element:

A. the processing element includes a memory having a plurality of storage locations identified by a series of memory addresses;

B. the message generator stores messages generated thereby in the storage locations such that messages to be transmitted to successive processing elements are stored in storage locations having successive memory addresses; and

C. the message transmitter retrieves messages from succeeding storage locations in the memory for transmission over the communications network, the message transmitters of processing elements which have successive network addresses selecting as initial storage locations those storage locations which have succeeding memory addresses.

8. A system as defined in claim 7 in which, in each processing element, the message transmitter includes:

A. a storage location identifier portion including;

i. a storage location identifier store for storing an address value that identifies a storage location whose message is to be transmitted;

ii. an initial storage location identification generator for generating an initial storage location identification value for storage in the storage location identifier store to be used to identify an initial storage location,

iii. a storage location identifier incrementation generator for incrementing the address value in the storage location identifier store;

B. a transmission portion for retrieving the message in the storage location identifier by the storage location identifier and transmitting the retrieved message over the communications network, and

C. a transmission controller for enabling;

i. the initial storage location identification generator to use the skew parameter value to generate the initial storage location identification value and to store the initial storage location identification value in the storage location identifier store, and

ii. the storage location identifier incrementation generator to increment the contents of the storage location identifier store; and

iii. the transmission portion to operate to transmit a message from the storage location identified by the contents of the storage location identifier store as identified by the initial storage location identification value as generated by the initial storage location identification generator in the storage location identifier store and each incrementation by the storage location identifier incrementation generator.

9. A system as defined in claim 8 in which, for each processing element, the storage location identifier incrementation generator increments the address value modulo a value corresponding to the number of messages generated by the processing element'"'"'s message generator.

10. A system as defined in claim 8 in which, in each processing element, the initial storage location identification generator generates the initial storage location identification value S₀ as S₀ =Up mod M, where "U" corresponds to the skew parameter value, "p" corresponds to the processing element'"'"'s network address, "M" corresponds to the number of messages generated by the processing element'"'"'s respective message generator, and "mod" refers to a "modulo" function.

11. A system as defined in claim 10 in which the initial storage location identification generator further generates the initial storage location identification value S₀ as S₀ =WUp mod M, where "W" is a roll parameter value.

12. A system as defined in claim 10 in which the initial storage location identification generator further generates the initial storage location identification value as a function of an initial offset value.

13. A system as defined in claim 12 in which the skew parameter value and the roll parameter value are both selected to be prime relative to the number of messages generated by each processing element.

14. A system as defined in claim 10 in which the transmission controller enables the storage location identifier incrementation generator to increment the contents of the storage location identifier store by an amount corresponding to the roll parameter value.

15. A system as defined in claim 14 in which, for each processing element, the storage location identifier incrementation generator, after incrementing contents of the storage location identifier store, generates the address value as the contents of the storage location identifier store modulo a value corresponding to the number of messages generated by the processing element'"'"'s message generator.

16. A system as defined in claim 1 further including a host for controlling the processing elements to operate in parallel.

5. A computer system as defined in clam 4 in which the mathematical function is S_t =S_t-1 +W (mod M), where S_t is related to the network address for the address portion associated with a message to be transmitted by the processing element, S_t-1 is related to the network address for the address portion associated with a message previously transmitted by the processing element, "W" represents the roll parameter value and "M" corresponds to the number of messages to be transmitted by the processing element during the message transfer operation, and "mod" refers to a modulo function.

17. A method for use in connection with a computer system comprising a plurality of processing elements interconnected by a communications network, the communications network having a series of network addresses each associated with one of said processing elements, the communications network transferring messages transmitted by the processing elements in accordance with an address portion associated with each message, each address portion identifying a network address, the communications network transferring each message to one of said processing elements whose network address is identified by the message'"'"'s address portion;

at least some of said processing elements during a message transfer operation generating a series of messages and transmitting them over the communications network, the method of controlling at least some of said processing elements to generate messages for transmission during a message transfer operation comprising the steps wherein each of said at least some of said processing elements;

A. transmits a first message whose address portion contains a network address which is a mathematical function of a skew parameter value and the processing element'"'"'s network address, and

B. after transmitting the first message, transmits successive messages having address portions containing successive network addresses following the first message'"'"'s network address in the series,

such that successive messages are associated with respective address portions which enable the communications network to transfer the messages to processing elements which have successive network addresses.

18. A method as defined in claim 17 in which the mathematical function is S₀ =Up mod M where S₀ is related to the network address for the address portion associated with the first message, "U" corresponds to the skew parameter value, "p" corresponds to the network address of the processing element generating the message, "M" corresponds to the number of messages to be transmitted by the processing element during the message transfer operation, and "mod" refers to a modulo function.

19. A method defined in claim 17 in which the series of network addresses has a network address range from a lower network address value to an upper network address value, with the series of network addresses sequencing from the lower network address value toward the upper network address value,

A. after successively transmitting messages respectively associated with address portions identifying network addresses in the portion of the network address range from the processing element'"'"'s network address toward the upper network address value, each processing element continuing to successively transmit messages respectively associated with address portions identifying network addresses in the portion of the network address range from the lower network address value toward the upper network address value.

20. A method as defined in claim 17 in which, for each processing element, the address portion associated with each message after said first message contains a network address which is a second mathematical function of a roll parameter value and a network address of the address portion associated with a message previously transmitted by said each processing element.

21. A method as defined in claim 20 in which the mathematical function is S_t=St-1 +W (mod M), is related to S_t represents the network address for the address portion associated with a message to be transmitted by the processing element, S_t-1 is related to the network address for the address portion associated with a message previously transmitted by the processing element, "W" represents the roll parameter value and "M" corresponds to the number of messages to be transmitted by the processing element during the message transfer operation, and "mod" refers to a modulo function.