Parallel computer system and program

• US 9,043,804 B2
• Filed: 04/02/2012
• Issued: 05/26/2015
• Est. Priority Date: 04/04/2011
• Status: Expired due to Fees

• Current Assignee: Hitachi, Ltd.
• Original Assignee: Hitachi, Ltd.
• Inventors: Hamamoto, Masaki, Yamada, Tetsuya, Tomoda, Atsushi, Miyamoto, Atsushi
• Primary Examiner(s): Bullock, Jr., Lewis A
• Assistant Examiner(s): Lu, Kevin X
• Application Number:
  US13/437,523
• Publication Number:
  US 20120254881A1
• Time in Patent Office: 1,149 Days
• US Class Current: 718/106
• CPC Class Codes: G06F 9/4887 involving deadlines, e.g. rate based, periodic
• CPC Class Codes: G06F 9/522 Barrier synchronisation

Abstract

There is provided a parallel computer system for performing barrier synchronization using a master node and a plurality of worker nodes based on the time to allow for an adaptive setting of the synchronization time. When a task process in a certain worker node has not been completed by a worker determination time, the particular worker node performs a communication to indicate that the process has not been completed, to a master node. When the communication has been received by a master determination time, the master node performs a communication to indicate that the process time is extended by a correction process time, in order to adjust and extend the synchronization time. In this way, it is possible to reduce the synchronization overhead associated with the execution of an application with a relatively large variation in the process time from a synchronization point to the next synchronization point.

Claims

1. A parallel computer system for performing parallel computation, the system comprising:

a plurality of computing units connected via a network, the computing units serve as a master node performing synchronization processes or as a worker node performing task processes, and one computing unit is designated as the master node and the other plurality of computing units are designated as the worker nodes,

wherein the master node;

sets a master determination time, before which the task processes in all the worker nodes are expected to be completed within a basic process time of each task process;

transmits a process start notification to the plurality of worker nodes;

checks whether a process-not-completed notification is received from any of the worker nodes at the master determination time;

if the process-not-completed notification is received, the master node transmits a process extension notification to the plurality of worker nodes regardless of whether the worker nodes have completed processing of the process tasks; and

if the process-not-completed notification is not received, the master node transmits a synchronization completion notification to the plurality of worker nodes;

wherein the worker nodes each;

set a worker determination time using the basic process time when the process start notification is received from the master node;

if the task process is not completed at the worker determination time, the worker node transmits the process-not-complete notification to the master node; and

if the task process is completed at the worker determination time, the worker node not sending the process-not-complete-notification and waits for the synchronization completion notification from the master node.

2. The parallel computer system according to claim 1,

wherein if the task process in at least one of the worker nodes has not been completed by the basic process time, the master node redefines the master determination time using a correction process time,

wherein if the process extension notification is received from the master node, the worker node redefines the worker determination time using the correction process time.

3. The parallel computer system according to claim 2,

wherein the master node transmits the synchronization completion notification or the process start notification, together with information of the basic process time,

wherein the master node transmits the process extension notification together with information of the correction process time.

4. The parallel computer system according to claim 2,

wherein the master node updates a value of the basic process time or correction process time using a predetermined calculation method.

5. The parallel computer system according to claim 1,

wherein each of the plurality of computing units comprise a plurality of communication control units, a processor, and a storage unit, the plurality of communication control units are connected to the processor, the storage unit, and the network.

6. The parallel computer system according to claim 5,

wherein one of the communication control units is used for transmitting and receiving the process start notification, the process-not-completed notification, the process extension notification, and the synchronization completion notification between the master node and the worker node.

7. The parallel computer system according to claim 5,

wherein the storage unit of the computing unit includes a memory area for storing master/worker information to identify whether the particular computing unit is the master node or the worker node.

8. The parallel computer system according to claim 7,

wherein the computing unit receives a program to be executed by the worker node from the computing unit serving as the master node, and stores the master/worker information in the memory area.

9. A program stored in a non-transitory computer readable medium and executed by a processor of a plurality of computing units connected via a network, the plurality of computing units are part of a parallel computer system for performing parallel computation the computing units each comprise the processor and the non-transitory computer readable medium, the computing units serve as a master node performing synchronization processes or as a worker node performing task processes, and one computing unit is designated as the master node and the other of the plurality of computing units are designated as the working nodes,

wherein the program causes the processor of the computing unit serving as the master node to perform the steps of:

setting a master determination time, before which the task processes in all the worker nodes are expected to be completed within a basic process time of each task process;

transmitting a process start notification to the plurality of worker nodes;

checking whether a process-not-completed notification is received from any of the worker nodes at the master determination time,

when the process-not-completed notification is received, transmitting a process extension notification to the plurality of worker nodes regardless of whether the worker nodes have completed processing of the process tasks; and

when the process-not-completed notification is not received, transmitting a synchronization completion notification to the plurality of worker nodes,

wherein the program causes a computing unit serving as a worker node to perform the steps of;

setting a worker determination time by using the basic process time when the process start notification is received from the master node;

when the task process is not completed at the worker determination time, transmitting the process-not-completed notification to the master node; and

when the task process is completed at the worker determination time, not sending the process-not-completed notification and waiting for the synchronization completion notification from the master node.

10. The program according to claim 9,

wherein when the task process in at least one of the worker nodes has not been completed by the basic process time, the program causes the processor of the computing unit serving as the master node to redefine the master determination time by using a correction process time,

wherein when the process extension notification is received from the master node, the program causes the processor of the computing unit serving as the worker node to redefine the worker determination time by using the correction process time.

11. The program according to claim 10,

wherein the program causes the processor of the computing unit serving as the master node to perform the steps of:

transmitting the synchronization completion notification or the process start notification, together with information of the basic process time; and

transmitting the process extension notification together with information of the correction process time.

12. The program according to claim 10,

wherein the program causes the processor of the computing unit serving as the master node to update a value of the basic process time or correction process time by a predetermined calculation method.

13. The program according to claim 9,

wherein the program causes the processor of the computing unit to store master/worker information in the non-transitory computer readable medium to identify whether the particular computing unit is the master node or the worker node.

14. The program according to claim 13,

wherein the program causes the processor of the computing unit serving as the master node to transmit a program to be executed by the worker node and an initial value of the program, to the computing unit serving as the worker node.

15. The program according to claim 14,

wherein when the program and the initial value of the program are received, the program causes the processor of the computing unit serving as the worker node to store the master/worker information indicating the worker node into the non-transitory computer readable medium.

16. A method for controlling a parallel computer system for performing parallel computation by connecting a plurality of computing units via a network, the computing units serve as a master node for performing synchronization processes or as a worker node for performing task processes, and one computing unit is designated as the master node while the other of the plurality of computing units are designated as the worker nodes, the method comprising the steps of:

setting, by the master node, a master determination time before which the task processes in all the worker nodes are expected to be completed within a basic process time of each task process;

transmitting, by the master node, a process start notification to the plurality of worker nodes;

checking, by the master node, whether a process-not-completed notification is received from any of the worker nodes at the master determination time,

when the process-not-completed notification is received, transmitting, by the master node, a process extension notification to the plurality of worker nodes regardless of whether the worker nodes have completed processing of the process tasks;

when the process-not-completed notification is not received, transmitting, by the master node, a synchronization completion notification to the plurality of worker nodes;

setting, by each worker node, a worker determination time by using the basic process time when the process start notification is received from the master node;

when the task process is not completed at the worker determination time, transmitting, by each worker node, the process-not-completed notification to the master node; and

when the task process is completed at the worker determination time, not sending, by the worker node, the process-not-completed notification and waiting, by each worker node, for the synchronization completion notification from the master node.