Method for decentralized load distribution in an event-driven system using localized migration between physically connected nodes and load exchange protocol preventing simultaneous migration of plurality of tasks to or from a same node

• US 8,479,216 B2
• Filed: 08/18/2009
• Issued: 07/02/2013
• Est. Priority Date: 08/18/2009
• Status: Active Grant

• Current Assignee: International Business Machines Corporation
• Original Assignee: International Business Machines Corporation
• Inventors: Lakshmanan, Geetika Tewari, Rabinovich, Yuri G., Schloss, Robert Jeffrey
• Primary Examiner(s): Bullock, Jr., Lewis A
• Assistant Examiner(s): Lu, Kevin X
• Application Number:
  US12/543,476
• Publication Number:
  US 20110047555A1
• Time in Patent Office: 1,414 Days
• Field of Search: 718/105
• US Class Current: 718/105
• CPC Class Codes: G06F 9/5088 involving task migration
• CPC Class Codes: Y02D 10/00 Energy efficient computing, e.g. low power processors, power management or thermal management

Abstract

A computer-implemented method, computer program product and computer readable storage medium directed to decentralized load distribution in an event-driven system. Included are receiving a data flow to be processed by a plurality of tasks at a plurality of nodes in the event-driven system having stateful and stateless event processing components, wherein the plurality of tasks are selected from the group consisting of hierarchical tasks (a task that is dependent on the output of another task), nonhierarchical tasks (a task that is not dependent on the output of another task) and mixtures thereof. Tasks are considered for migration to distribute the system load of processing tasks. The target node, to which the at least one target task is migrated, is chosen wherein the target node meets predetermined criteria in terms of load distribution quality. The computer-implemented method, computer program product and computer readable storage medium of the present invention may also include migrating tasks to target nodes to reduce cooling costs and selecting at least one node to go into quiescent mode.

Claims

1. A method of decentralized load distribution in an event-driven system, the method comprising the steps of:

receiving a data flow to be processed by a plurality of tasks at a plurality of nodes in the event- driven system having stateful and stateless event processing components wherein the plurality of tasks are selected from the group consisting of hierarchical tasks, wherein a hierarchical task is a task that is dependent on the output of another task, nonhierarchical tasks, wherein a nonhierarchical task is a task that is not dependent on the output of another task, and mixtures thereof;

collecting statistics about the execution of each task hosted at each node;

creating a list of neighbor nodes, using the collected statistics, to which a task can be partially or wholly transferred;

selecting at least one target task at a source node for consideration to migrate, to a target node, the target node selected from the list of neighbor nodes, to distribute the system load of processing the at least one target task;

choosing the target node to which the at least one target task can be migrated wherein the target node meets predetermined criteria in terms of load distribution quality such that the target node must be physically connected to;

(a) the source node hosting the target task, (b) a parent node hosting a parent task of the target task, and (c) a child node hosting a child task of the target task; and

establishing a load exchange protocol at each node for governing the number of migrations of target tasks,

wherein local decentralized load migrations lead to overall system load distribution in the event-driven system; and

wherein the load exchange protocol comprises all of the following;

(i) a decision to migrate the target task should not lead to oscillation such that a target task is migrated more than once to the target node and back to a node first hosting the target task;

(ii) no simultaneous migration of 2 or more tasks to a single target node in a single machine cycle shall occur;

(iii) no simultaneous migrations of 2 or more tasks from a node in a single machine cycle shall occur; and

(iv) an end result of target task migration should improve load distribution in some manner.

2. The method of claim 1 further comprising splitting the at least one target task into 2 or more target subtasks.

3. The method of claim 1 wherein the predetermined criteria in terms of load distribution quality further comprises:

(i) the target node must have sufficient available load and memory to host the target task;

(ii) minimizing the number of physical links in the data flow which passes through the target task;

(iii) eliminating cycles;

(iv) the target node should have negative load correlation coefficient; and

(v) post-migration utilization of the node hosting the target task and the target node are less than a predetermined maximum threshold value.

4. The method of claim 1 wherein each node has all of the following load exchange constraints:

(i) during a first predetermined period of time, all target tasks are migrated either upstream or downstream and after a second predetermined period of time, the direction of target task migration may be switched;

(ii) if a target task is migrated to a target node, then any nearest neighbor downstream target tasks prior to the migration cannot be migrated for a predetermined amount of time;

(iii) after migration to a target node, a target task must remain at the target node for a predetermined amount of time; and

(iv) two nodes cannot migrate target tasks to a target node at the same time.

5. The method of claim 1 further comprising the steps of:

migrating at least one target task to a target node and executing the at least one target task on the target node; and

collecting statistics about the execution of the at least one target task on the target node.

6. The method of claim 1 further comprising the step of selecting at least one node to transition into quiescent mode such that the at least one node migrates its tasks to a recipient node before becoming quiescent, the recipient node being physically connected to:

(a) the at least one node, (b) any nodes from which the at least one node receives input data and (c) any nodes to which the at least one node sends output data.

7. The method of claim 1 wherein the step of collecting statistics includes collecting statistics about the characteristics of the node including its heat characteristics, wherein the node hosting the target task has a first temperature and the target node has a second temperature and further comprising the step of migrating the target task to the target node provided the first temperature is lowered and the second temperature is below a predetermined acceptable heat threshold.

8. A computer program product comprising:

a nontransitory computer recordable storage medium having computer readable program code for decentralized load distribution in an event-driven system comprising;

computer readable program code configured for receiving a data flow to be processed by a plurality of tasks at a plurality of nodes in the event-driven system having stateful and stateless event processing components, wherein the plurality of tasks are selected from the group consisting of hierarchical tasks, wherein a hierarchical task is a task that is dependent on the output of another task, nonhierarchical tasks, wherein a nonhierarchical task is a task that is not dependent on the output of another task, and mixtures thereof;

computer readable program code configured for collecting statistics about each task hosted at each node;

computer readable program code configured for creating a list of neighbor nodes, using the collected statistics, to which a task can be partially or wholly transferred;

computer readable program code configured for selecting at least one target task at a source node for consideration to migrate, to a target node, the target node selected from the list of neighbor nodes, to distribute the system load of processing the at least one target task;

computer readable program code configured the target node to which the at least one target task can be migrated wherein the target node meets predetermined criteria in terms of load distribution quality such that the target node must be physically connected to;

(a) the source node hosting the target task, (b) a parent node hosting a parent task of the target task, and (c) a child node hosting a child task of the target task; and

computer readable program code configured for establishing a load exchange protocol at each node for governing the number of migrations of target tasks,

wherein decentralized load migrations lead to overall system load distribution in the event-driven system; and

wherein the load exchange protocol comprises all of the following;

(i) a decision to migrate the target task should not lead to oscillation such that a target task is migrated more than once to the target node and back to a node first hosting the target task;

(ii) no simultaneous migration of 2 or more tasks to a single target node in a single machine cycle shall occur;

(iii) no simultaneous migrations of 2 or more tasks from a node in a single machine cycle shall occur; and

(iv) an end result of target task migration should improve load distribution in some manner.

9. The computer program product of claim 8 further comprising computer readable program code configured for splitting the at least one target task into 2 or more target subtasks.

10. The computer program product of claim 8 wherein the predetermined criteria in terms of load distribution quality further comprises:

(i) the target node must have sufficient available load and memory to host the target task;

(ii) minimizing the number of physical links in the data flow which passes through the target task;

(iii) eliminating cycles;

(iv) the target node should have negative load correlation coefficient; and

(v) post-migration utilization of the node hosting the target task and the target node are less than a predetermined maximum threshold value.

11. The computer program product of claim 8 wherein each node has all of the following load exchange constraints:

(i) during a first predetermined period of time, all target tasks are migrated either upstream or downstream and after a second predetermined period of time, the direction of target task migration may be switched;

(ii) if a target task is migrated to a target node, then any nearest neighbor downstream target tasks prior to the migration cannot be migrated for a predetermined amount of time;

(iii) after migration to a target node, a target task must remain at the target node for a predetermined amount of time specifically set for that node; and

(iv) two nodes cannot migrate target tasks to a target node at the same time.

12. The computer program product of claim 8 further comprising:

computer readable program code configured for migrating at least one target task to a target node and executing the at least one target task on the target node; and

computer readable program code configured for collecting statistics about the execution of the at least one target task on the target node.

13. The computer program product of claim 8 further comprising computer readable program code configured for selecting at least one node to transition into quiescent mode such that the at least one node migrates its tasks to a recipient node before becoming quiescent, the recipient node being physically connected to:

(a) the at least one node, (b) any nodes from which the at least one node receives input data and (c) any nodes to which the at least one node sends output data.

14. The computer program product of claim 8 wherein the computer usable program code configured for collecting statistics includes collecting statistics about the characteristics of the node including its heat characteristics, wherein the node hosting the target task has a first temperature and the target node has a second temperature and further comprising computer usable program code configured for migrating the target task to the target node provided the first temperature is lowered and the second temperature is below a predetermined acceptable heat threshold.

15. A nontransitory computer readable storage medium storing instructions that, when executed by a computer, causes the computer to perform a method of decentralized load distribution in an event-driven system, the method comprising the steps of:

receiving a data flow to be processed by a plurality of tasks at a plurality of nodes in the event- driven system having stateful and stateless event processing components, wherein the plurality of tasks are selected from the group consisting of hierarchical tasks, wherein a hierarchical task is a task that is dependent on the output of another task, nonhierarchical tasks, wherein a nonhierarchical task is a task that is not dependent on the output of another task, and mixtures thereof;

collecting statistics about each task hosted at each node;

creating a list of neighbor nodes, using the collected statistics, to which a task can be partially or wholly transferred;

selecting at least one target task at a source node for consideration to migrate, to a target node, the target node selected from the list of neighbor nodes, to distribute the system load of processing the at least one target task;

choosing the target node to which the at least one target task can be migrated wherein the target node meets predetermined criteria in terms of load distribution quality such that the target node must be physically connected to;

(a) the source node hosting the target task, (b) a parent node hosting a parent task of the target task, and (c) a child node hosting a child task of the target task; and

establishing a load exchange protocol at each node for governing the number of migrations of target tasks,

wherein local decentralized load migrations lead to overall system load distribution in the event-driven system; and

wherein the load exchange protocol comprises all of the following;

(i) a decision to migrate the target task should not lead to oscillation such that a target task is migrated more than once to the target node and back to a node first hosting the target task;

(ii) no simultaneous migration of 2 or more tasks to a single target node in a single machine cycle shall occur;

(iii) no simultaneous migrations of 2 or more tasks from a node in a single machine cycle shall occur; and

(iv) an end result of target task migration should improve load distribution in some manner.

16. The computer readable storage medium of claim 15 further comprising splitting the at least one target task into 2 or more target subtasks.

17. The computer readable storage medium of claim 15 wherein the predetermined criteria in terms of load distribution quality further comprises:

(i) the target node must have sufficient available load and memory to host the target task;

(ii) minimizing the number of physical links in the data flow which passes through the target task;

(iii) eliminating cycles;

(iv) the target node should have negative load correlation coefficient; and

(v) post-migration utilization of the node hosting the target task and the target node are less than a predetermined maximum threshold value.

18. The computer readable storage medium of claim 15 wherein each node has all of the following load exchange constraints:

(i) during a first predetermined period of time, all target tasks are migrated either upstream or downstream and after a second predetermined period of time, the direction of target task migration may be switched;

(ii) if a target task is migrated to a target node, then any nearest neighbor downstream target tasks prior to the migration cannot be migrated for a predetermined amount of time;

(iii) after migration to a target node, a target task must remain at the target node for a predetermined amount of time; and

(iv) two nodes cannot migrate target tasks to a target node at the same time.

19. The computer readable storage medium of claim 15 further comprising the steps of:

migrating at least one target task to a target node and executing the at least one target task on the target node; and

collecting statistics about the execution of the at least one target task on the target node.

20. The computer readable storage medium of claim 15 further comprising the step of selecting at least one node to transition into quiescent mode such that the at least one node migrates its tasks to a recipient node before becoming quiescent, the recipient node being physically connected to:

(a) the at least one node, (b) any nodes from which the at least one node receives input data and (c) any nodes to which the at least one node sends output data.

21. The computer readable storage medium of claim 15 wherein the step of collecting statistics includes collecting statistics about the characteristics of the node including its heat characteristics, wherein the node hosting the target task has a first temperature and the target node has a second temperature and further comprising the step of migrating the target task to the target node provided the first temperature is lowered and the second temperature is below a predetermined acceptable heat threshold.