Parallel program execution time with message consolidation

• US 5,781,787 A
• Filed: 04/21/1995
• Issued: 07/14/1998
• Est. Priority Date: 04/21/1995
• Status: Expired due to Fees

• Current Assignee: Loral Federal Systems Company (Martin Marietta Corporation)
• Original Assignee: Lockheed Martin Corporation (Martin Marietta Corporation)
• Inventors: Ghose, Kanad, Shafer, Stephen Ray
• Primary Examiner(s): Coleman, Eric
• Application Number:
  US08/426,805
• Time in Patent Office: 1,180 Days
• Field of Search: 395/800, 395/566
• US Class Current: 712/28
• CPC Class Codes: G06F 9/524 Deadlock detection or avoidance
• CPC Class Codes: G06F 9/546 Message passing systems or structures, e.g. queues

Abstract

In distributed memory multiprocessors, communication between processing elements (PEs) can have a significant impact on the overall computation time. In addition, contention for the communication links can often make PEs wait even longer for a message than would normally be required. Because of this, it is important to minimize the effects of inter-processor communication time. The present invention reduces the execution time of a parallel program by merging messages (also called message combining, or message consolidation) after the program has already been partitioned and scheduled onto the PEs. The data from two (or more) messages are combined and sent in a single communication by locating sections of the system where a merge will affect the overall execution time, and determining before the merge takes place whether it will positively or negatively affect the system.

Claims

1. A method of determining which messages in a program for a distributed memory parallel processor system can be merged without causing deadlock of the system when the program is run, comprising the steps of:

(a) determining regions in the code allocated to each processor, where new regions begin at the receive statements of entering messages;

(b) identifying two consecutive outgoing messages to a second processor and sent from a first processor, where outgoing messages will be consecutive when considering only messages going to the second processor;

(c) checking if the two consecutive messages are crossed, with crossed messages having a first receive statement whose message is sent first from the first processor, said first receiving statement being located after the receive statement which is to receive the outgoing message from the second consecutive outgoing message; and

(d) determining for non-crossed messages whether there is a chain that originates between the two receive statements and terminates between the two send statements for the consecutive messages being considered for merging and not merging the messages if such a path is found, thereby avoiding deadlock.

2. A method carried out on a computer of merging two or more messages of a program for a distributed memory parallel processor system, so that the messages can be sent as one message to reduce the overall computation time and bus loading of the system while running the program, after the program that has been scheduled, comprising the steps of:

(a) determining regions in the program code allocated to each processor, where new regions begin at the receive statements of entering messages;

(b) estimating start time of each region based on whether it takes longer for the execution of the previous region to be completed than for the message to arrive which defines the beginning of the region, or if it takes longer for the message to arrive which defines the beginning of the region than the time the previous region takes to complete its execution;

(c) identifying two consecutive outgoing messages from a first processor to a second processor, where outgoing messages will be consecutive when considering only messages going to the second processor;

(d) checking if the two consecutive messages are crossed, with crossed messages having a first receive statement whose message is sent first from the first processor, the first receive statement being located after the receive statement which is to receive the outgoing message from the second consecutive outgoing message; and

(e) determining for non-crossed messages whether there is a chain that originates between the two receive statements and terminates between the two send statements for the consecutive messages being considered for merging and not merging the messages if such a path is found, thereby avoiding deadlock.

3. The method of claim 2, further comprising the steps of:

determining for non-crossed messages that would not produce a deadlock whether the start time of the portion of the program waiting for the combined receive statement message is delayed;

modifying the program by merging the consecutive messages if no delay occurs.

4. The method of claim 3 further comprising the steps of:

determining the difference in execution time with and without the merge which caused a delay of the start time; and

modifying the program by performing the merge if the execution time was not increased.

5. The method of claim 2 further comprising the steps of:

determining for crossed messages the time it takes to send a first message compared to the flat overhead associated with a send statement; and

modifying the program by performing the merge if the time to send the message is less than or equal to the flat overhead.

6. The method of claim 5, further comprising the steps of:

determining for crossed messages where the time to send the message is greater than the flat overhead, whether the change in arrival time at the first receive statement before and after the merge, delays the start time for the portion of the program waiting for the receive statement message;

modifying the program by performing the merge if no delay occurs.

7. The method of claim 6, further comprising the steps of:

determining the difference in execution time with and without the merge which caused a delay of the start time; and

modifying the program by performing the merge if the execution time was not increased.