IP Library › Granted Patent US 10,193,969
Granted Patent B2
US 10,193,969 · App. 15/137,221 · Granted Jan 29, 2019

Parallel processing system, method, and storage medium

Inventors: Masahiro Miwa (Kawaguchi, JP); Kohta Nakashima (Kawasaki, JP)
Assignee: FUJITSU LIMITED
H04L67/1044H04L67/10
View Patent ↗
Loading inventors, assignments & file history…
Monitor This Case
Get email alerts when status or documents change.
Order Certified Copies
Most orders are placed with the USPTO same day — all within 24 business hours.
Order via The Patent Place →
Pre-filled with this patent's details
Quick Facts
Patent No.
US 10,193,969
App. No.
15/137,221
Granted
Jan 29, 2019
Kind
B2
Abstract

A system is a multi-layered fullmesh system in which layers of fullmesh systems, having Leaf switches fullmesh-coupled to each other, are coupled to each other, the system including nodes and performing applications, at least one of nodes being coupled to each of the Leaf switches. The parallel processing system includes circuitry configured to: compare communication recording information in which a number of times of communication between nodes during execution of an application is recorded with communication pattern information in which assignment information indicating which nodes having an intra-layer or inter-layer connection relationship with a Leaf switch are assigned is specified for each communication pattern; extract assignment information of a communication pattern which is the most similar to communication indicated by the communication recording information, based on the communication pattern information; and assign nodes that subsequently execute the application, based on the extracted assignment information.

Claims (25)

1. A parallel processing system which is a multi-layered fullmesh system in which a plurality of layers of fullmesh systems, having a plurality of Leaf switches fullmesh-coupled to each other, are coupled to each other, the parallel processing system including a plurality of nodes, the parallel processing system being configured to perform a parallel arithmetic operation of applications, at least one of the plurality of nodes being coupled to each of the plurality of Leaf switches, the parallel processing system comprising:

circuitry configured to:

obtain communication recording information in which a number of times of communication between the plurality of nodes during execution of an application is recorded;

obtain communication pattern information in which assignment information indicating which connection topology of a fullmesh topology and a fattree topology is to be selected for each a plurality of communication patterns, in the fullmesh topology, the plurality of nodes having an intra-layer connection relationship, and in the fattree topology, the plurality of nodes having inter-layer connection relationship;

select a first communication pattern from the plurality of communication patterns based on the number of times of communication;

identify, based on the communication pattern information, a connection topology from the fullmesh topology and the fattree topology corresponding to the selected first communication pattern; and

assign, based on the identified connection topology, nodes included in the parallel processing system that execute the application.

2. The parallel processing system according to claim 1 , wherein the circuitry is configured to:

select the first communication pattern which is the most similar to communication indicated by communication recording information at the time of first executing the application, based on the communication pattern information.

3. A method executed by circuitry in a parallel processing system which is a multi-layered fullmesh system in which a plurality of layers of fullmesh systems, having a plurality of Leaf switches fullmesh-coupled to each other, are coupled to each other, the parallel processing system including a plurality of nodes, the parallel processing system being configured to perform a parallel arithmetic operation of applications, at least one of the plurality of nodes being coupled to each of the plurality of Leaf switches, the method comprising:

obtaining communication recording information in which a number of times of communication between the plurality of nodes during execution of an application is recorded;

obtaining communication pattern information in which assignment information indicating which connection topology of a fullmesh topology and a fattree topology is to be selected for each a plurality of communication patterns, in the fullmesh topology, the plurality of nodes having an intra-layer connection relationship, and in the fattree topology, the plurality of nodes having inter-layer connection relationship;

selecting a first communication pattern from the plurality of communication patterns based on the number of times of communication;

identifying, based on the communication pattern information, a connection topology from the fullmesh topology and the fattree topology corresponding to the selected first communication pattern; and

assigning, based on the identified connection topology, nodes included in the parallel processing system that execute the application.

4. The method according to claim 3 , further comprising:

selecting the first communication pattern which is the most similar to communication indicated by communication recording information at the time of first executing the application, based on the communication pattern information.

5. A non-transitory storage medium storing a program for causing circuitry in a parallel processing system to execute a process, the parallel processing system being a multi-layered fullmesh system in which a plurality of layers of fullmesh systems, having a plurality of Leaf switches fullmesh-coupled to each other, are coupled to each other, the parallel processing system including a plurality of nodes, the parallel processing system being configured to perform a parallel arithmetic operation of applications, at least one of the plurality of nodes being coupled to each of the plurality of Leaf switches, the process comprising:

obtaining communication recording information in which a number of times of communication between the plurality of nodes during execution of an application is recorded;

obtaining communication pattern information in which assignment information indicating which connection topology of a fullmesh topology and a fattree topology is to be selected for each a plurality of communication patterns, in the fullmesh topology, the plurality of nodes having an intra-layer connection relationship, and the fattree topology corresponding to the selected first communication pattern inter-layer connection relationship;

selecting a first communication pattern from the plurality of communication patterns based on the number of times of communication;

identifying, based on the communication pattern information, a connection topology from the fullmesh topology and the fattree topology corresponding to the selected first communication pattern; and

assigning, based on the identified connection topology, nodes included in the parallel processing system that execute the application.

6. The storage medium according to claim 5 , wherein the process further comprising:

selecting the first communication pattern which is the most similar to communication indicated by communication recording information at the time of first executing the application, based on the communication pattern information.

Assignments (1)
ASSIGNMENT OF ASSIGNOR'S INTEREST Recorded Apr 28, 2016
From: MIWA, MASAHIRO; NAKASHIMA, KOHTA
To: FUJITSU LIMITED
Reel/Frame 038410/0017 →
Priority Claims (1)
JP 2015-111287 · Jun 1, 2015 · national
Continuity (1)
Related Publication 20160352824A1 · Dec 1, 2016