Administering connection identifiers for collective operations in a parallel computer

• US 9,053,226 B2
• Filed: 03/04/2013
• Issued: 06/09/2015
• Est. Priority Date: 07/30/2010
• Status: Expired due to Fees

• Current Assignee: International Business Machines Corporation
• Original Assignee: International Business Machines Corporation
• Inventors: Faraj, Daniel A., Smith, Brian E.
• Primary Examiner(s): NANO, SARGON N
• Application Number:
  US13/783,713
• Publication Number:
  US 20130179620A1
• Time in Patent Office: 827 Days
• Field of Search: 709201-205, 709/212, 709/213, 709/216, 709/217, 709/234, 709/237, 709/250, 712/28, 712/225
• US Class Current: 1/1
• CPC Class Codes: G06F 13/36 for access to common bus or bus system
• CPC Class Codes: G06F 9/526 Mutual exclusion algorithms

Abstract

Administering connection identifiers for collective operations in a parallel computer, including prior to calling a collective operation, determining, by a first compute node of a communicator to receive an instruction to execute the collective operation, whether a value stored in a global connection identifier utilization buffer exceeds a predetermined threshold; if the value stored in the global ConnID utilization buffer does not exceed the predetermined threshold: calling the collective operation with a next available ConnID including retrieving, from an element of a ConnID buffer, the next available ConnID and locking the element of the ConnID buffer from access by other compute nodes; and if the value stored in the global ConnID utilization buffer exceeds the predetermined threshold: repeatedly determining whether the value stored in the global ConnID utilization buffer exceeds the predetermined threshold until the value stored in the global ConnID utilization buffer does not exceed the predetermined threshold.

Claims

1. A method of administering connection identifiers for collective operations in a parallel computer, the method comprising:

prior to calling a collective operation, if a value representing a number of connection identifiers in use and stored in a global connection identifier (‘

ConnID’

) utilization buffer does not exceed a predetermined threshold,;

calling the collective operation with a next available ConnID including, atomically;

retrieving, from an element of a ConnID buffer, the next available ConnID and locking the element of the ConnID buffer from access by other compute nodes; and

if the value stored in the global ConnID utilization buffer exceeds the predetermined threshold;

repeatedly determining whether the value stored in the global ConnID utilization buffer exceeds the predetermined threshold until the value stored in the global ConnID utilization buffer does not exceed the predetermined threshold.

2. The method of claim 1 wherein determining whether the value stored in the global ConnID utilization buffer exceeds the predetermined threshold further comprises:

atomically fetching, by a DMA engine of the first compute node, the value stored in the global ConnID utilization buffer and incrementing the stored value; and

determining whether the fetched value exceeds the predetermined threshold.

3. The method of claim 2 further comprising:

upon completion of the collective operation, atomically;

unlocking the element of the ConnID buffer storing the retrieved ConnID; and

decrementing the value stored in the global ConnID utilization buffer.

4. The method of claim 1 wherein calling the collective operation with the next available ConnID further comprises:

placing, by a DMA engine of the first compute node, in a predefined memory location in all other nodes of the communicator a value representing an instruction to wait for a ConnID; and

upon retrieving the next available ConnID, placing, by the DMA engine, in predefined memory location in all other nodes of the communicator, the retrieved ConnID; and

upon completion of the collective operation, the method further comprises clearing, from the predefined memory location in all the other nodes of the communicator, the retrieved ConnID.

5. The method of claim 1 wherein calling the collective operation with the next available ConnID further comprises:

determining, by the first node, whether a ConnID is stored in a predefined memory location of a master node of the communicator;

atomically;

retrieving, from the element of the ConnID buffer, the next available ConnID and locking the element of the ConnID buffer from access by other compute nodes only if no ConnID is stored in the predefined memory location of the master node;

again determining, after retrieving the next available ConnID, whether a ConnID is stored in the predefined memory location of the master node;

if there is a ConnID stored in the predefined memory location of the master node after retrieving the next available ConnID, atomically unlocking the element of the ConnID buffer;

if there is no ConnID stored in the predefined memory location of the master node after retrieving the next available ConnID;

placing, by a DMA engine of the first node, the retrieved ConnID in the predefined memory location of the master node; and

upon completion of the collective operation, the method further comprises clearing the retrieved ConnID from the predefined memory location of the master node.

6. The method of claim 1 wherein retrieving, from an element of a ConnID buffer, a next available ConnID further comprises incrementing a ConnID buffer pointer to a next, unlocked element of the ConnID buffer.

7. An apparatus for administering connection identifiers for collective operations in a parallel computer, the parallel computer having a plurality of compute nodes organized into a communicator, the apparatus comprising a computer processor and a computer memory operatively coupled to the computer processor, the computer memory having disposed within it computer program instructions capable of:

prior to calling a collective operation, if a value representing a number of connection identifiers in use and stored in a global connection identifier (‘

ConnID’

) utilization buffer does not exceed a predetermined threshold,;

calling the collective operation with a next available ConnID including, atomically;

retrieving, from an element of a ConnID buffer, the next available ConnID and locking the element of the ConnID buffer from access by other compute nodes; and

if the value stored in the global ConnID utilization buffer exceeds the predetermined threshold;

repeatedly determining whether the value stored in the global ConnID utilization buffer exceeds the predetermined threshold until the value stored in the global ConnID utilization buffer does not exceed the predetermined threshold.

8. The apparatus of claim 7 wherein determining whether the value stored in the global ConnID utilization buffer exceeds the predetermined threshold further comprises:

atomically fetching, by a DMA engine of the first compute node, the value stored in the global ConnID utilization buffer and incrementing the stored value; and

determining whether the fetched value exceeds the predetermined threshold.

9. The apparatus of claim 8 further comprising:

upon completion of the collective operation, atomically;

unlocking the element of the ConnID buffer storing the retrieved ConnID; and

decrementing the value stored in the global ConnID utilization buffer.

10. The apparatus of claim 7 wherein calling the collective operation with the next available ConnID further comprises:

placing, by a DMA engine of the first compute node, in a predefined memory location in all other nodes of the communicator a value representing an instruction to wait for a ConnID; and

upon retrieving the next available ConnID, placing, by the DMA engine, in predefined memory location in all other nodes of the communicator, the retrieved ConnID; and

upon completion of the collective operation, the apparatus further comprises computer program instructions capable of clearing, from the predefined memory location in all the other nodes of the communicator, the retrieved ConnID.

11. The apparatus of claim 7 wherein calling the collective operation with the next available ConnID further comprises:

determining, by the first node, whether a ConnID is stored in a predefined memory location of a master node of the communicator;

atomically;

retrieving, from the element of the ConnID buffer, the next available ConnID and locking the element of the ConnID buffer from access by other compute nodes only if no ConnID is stored in the predefined memory location of the master node;

again determining, after retrieving the next available ConnID, whether a ConnID is stored in the predefined memory location of the master node;

if there is a ConnID stored in the predefined memory location of the master node after retrieving the next available ConnID, atomically unlocking the element of the ConnID buffer;

if there is no ConnID stored in the predefined memory location of the master node after retrieving the next available ConnID;

placing, by a DMA engine of the first node, the retrieved ConnID in the predefined memory location of the master node; and

upon completion of the collective operation, the apparatus further comprises computer program instructions capable of clearing the retrieved ConnID from the predefined memory location of the master node.

12. The apparatus of claim 7 wherein retrieving, from an element of a ConnID buffer, a next available ConnID further comprises incrementing a ConnID buffer pointer to a next, unlocked element of the ConnID buffer.

13. A computer program product for administering connection identifiers for collective operations in a parallel computer, the parallel computer having a plurality of compute nodes organized into a communicator, the computer program product disposed in a non-transitory computer readable storage medium, the computer program product comprising computer program instructions capable of:

prior to calling a collective operation, if a value representing a number of connection identifiers in use and stored in a global connection identifier (‘

ConnID’

) utilization buffer does not exceed a predetermined threshold,;

calling the collective operation with a next available ConnID including, atomically;

retrieving, from an element of a ConnID buffer, the next available ConnID and locking the element of the ConnID buffer from access by other compute nodes; and

if the value stored in the global ConnID utilization buffer exceeds the predetermined threshold;

repeatedly determining whether the value stored in the global ConnID utilization buffer exceeds the predetermined threshold until the value stored in the global ConnID utilization buffer does not exceed the predetermined threshold.

14. The computer program product of claim 13 wherein determining whether the value stored in the global ConnID utilization buffer exceeds the predetermined threshold further comprises:

atomically fetching, by a DMA engine of the first compute node, the value stored in the global ConnID utilization buffer and incrementing the stored value; and

determining whether the fetched value exceeds the predetermined threshold.

15. The computer program product of claim 14 further comprising:

upon completion of the collective operation, atomically;

unlocking the element of the ConnID buffer storing the retrieved ConnID; and

decrementing the value stored in the global ConnID utilization buffer.

16. The computer program product of claim 13 wherein calling the collective operation with the next available ConnID further comprises:

placing, by a DMA engine of the first compute node, in a predefined memory location in all other nodes of the communicator a value representing an instruction to wait for a ConnID; and

upon retrieving the next available ConnID, placing, by the DMA engine, in predefined memory location in all other nodes of the communicator, the retrieved ConnID; and

upon completion of the collective operation, the computer program product further comprises computer program instructions capable of clearing, from the predefined memory location in all the other nodes of the communicator, the retrieved ConnID.

17. The computer program product of claim 13 wherein calling the collective operation with the next available ConnID further comprises:

determining, by the first node, whether a ConnID is stored in a predefined memory location of a master node of the communicator;

atomically;

retrieving, from the element of the ConnID buffer, the next available ConnID and locking the element of the ConnID buffer from access by other compute nodes only if no ConnID is stored in the predefined memory location of the master node;

again determining, after retrieving the next available ConnID, whether a ConnID is stored in the predefined memory location of the master node;

if there is a ConnID stored in the predefined memory location of the master node after retrieving the next available ConnID, atomically unlocking the element of the ConnID buffer;

if there is no ConnID stored in the predefined memory location of the master node after retrieving the next available ConnID;

placing, by a DMA engine of the first node, the retrieved ConnID in the predefined memory location of the master node; and

upon completion of the collective operation, the computer program product further comprises computer program instructions capable of clearing the retrieved ConnID from the predefined memory location of the master node.

18. The computer program product of claim 13 wherein retrieving, from an element of a ConnID buffer, a next available ConnID further comprises incrementing a ConnID buffer pointer to a next, unlocked element of the ConnID buffer.