Search Images Maps Play YouTube News Gmail Drive More »
Sign in
Screen reader users: click this link for accessible mode. Accessible mode has the same essential features but works better with your reader.

Patents

  1. Advanced Patent Search
Publication numberUS20060112388 A1
Publication typeApplication
Application numberUS 10/994,852
Publication dateMay 25, 2006
Filing dateNov 22, 2004
Priority dateNov 22, 2004
Also published asUS8185908, US20080276242
Publication number10994852, 994852, US 2006/0112388 A1, US 2006/112388 A1, US 20060112388 A1, US 20060112388A1, US 2006112388 A1, US 2006112388A1, US-A1-20060112388, US-A1-2006112388, US2006/0112388A1, US2006/112388A1, US20060112388 A1, US20060112388A1, US2006112388 A1, US2006112388A1
InventorsMasaaki Taniguchi, Harunobu Kubo
Original AssigneeMasaaki Taniguchi, Harunobu Kubo
Export CitationBiBTeX, EndNote, RefMan
External Links: USPTO, USPTO Assignment, Espacenet
Method for dynamic scheduling in a distributed environment
US 20060112388 A1
Abstract
A method and system is provided for assigning programs in a workflow to one or more nodes for execution. Prior to the assignment, a priority of execution of each program is calculated in relation to its dependency upon data received and transmitted data. Based upon the calculated priority and the state of each of the nodes, the programs in the workflow are dynamically assigned to one or more nodes for execution. In addition to the node assignment based upon priority, preemptive execution of the programs in the workflow is determined so that the programs in the workflow may not preemptively be executed at a selected node in response to the determination.
Images(9)
Previous page
Next page
Claims(20)
1. A method for assigning resources to a plurality of processing nodes comprising:
deciding priority of execution dependency of a program;
dynamically assigning said program to a node based upon said priority and in accordance with a state of each node in a multinode system;
determining preemptive execution of said program; and
executing said program at a designated node non-preemptively in response to a positive determination.
2. The method of claim 1, wherein said priority is based upon a criteria selected from a group consisting of: topological sorting method, and a shortest path length from a start program.
3. The method of claim 1, wherein the step of deciding priority of execution dependency of a program includes normalizing said priority.
4. The method of claim 1, wherein the step of deciding priority of execution dependency of a program includes assigning said program to a logical node based upon an estimated computation and transmission cost.
5. The method of claim 4, further comprising storing said logical node assignment in a workflow database of a global scheduler based upon said estimated costs.
6. The method of claim 1, wherein the step of dynamically assigning said program to a node includes assigning said program to a physical node at time of execution.
7. The method of claim 1, wherein the step of dynamically assigning said program to a node includes assigning said program to a logical node based upon estimated computation and transmission costs.
8. A system comprising:
a plurality of operating nodes;
a scheduling manager adapted to decide priority of execution dependency of a program;
a global scheduler adapted to dynamically assign said program to a node based upon said priority and a state of each node in said system; and
a program manager adapted to determine applicability of preemptive execution of said program, and to non-preemptively execute said program at a designated node in response to a positive determination.
9. The system of claim 8, wherein said priority is based upon a criteria selected from a group consisting of: topological sorting method, and a shortest path length from a start program.
10. The system of claim 8, wherein said scheduling manager is adapted to normalize said priority.
11. The system of claim 8, wherein said scheduling manager is adapted to assign said program to a logical node based upon estimated computation and transmission costs.
12. The system of claim 11, further comprising a workflow database adapted to store said logical node assignment based upon said estimated costs.
13. The system of claim 8, wherein said global scheduler is adapted to assign said program to a physical node at time of execution.
14. An article comprising:
a computer-readable signal-bearing medium;
a plurality of operating nodes in said medium;
means in said medium for deciding priority of execution dependency of a program;
means in said medium for dynamically assigning said program to one of said nodes based upon said priority and a state of each node in said system; and
means in said medium for determining applicability of preemptive execution of said program, and to non-preemptively execute said program at a designated node in response to a positive determination.
15. The article of claim 14, wherein said medium is selected from a group consisting of: a recordable data storage medium, and a modulated carrier signal.
16. The article of claim 14, wherein said means for deciding priority of execution dependency includes criteria selected from a group consisting of: topological sorting method, and a shortest path length from a start program.
17. The article of claim 14, wherein said means for dynamically assigning said program to a node based upon said priority and a state of each node in said system normalizes priority of execution.
18. The article of claim 14, wherein said means for deciding priority of execution dependency of a program includes assigning said program to a node based upon said priority and a state of each node in said system assigns said program to a logical node based upon estimated computation and transmission costs.
19. The article claim 18, further comprising means in the medium for storing said logical node assignment based upon said estimated costs.
20. The article of claim 14, wherein said means for dynamically assigning said program to one of said node based upon said priority and state of each node includes assigning said program to a physical node at time of execution.
Description
BACKGROUND OF THE INVENTION

1. Technical Field

This invention relates to a method and system for dynamically scheduling programs for execution on one or more nodes.

2. Description of the Prior Art

A directed acyclic graph (DAG) includes a set of nodes connected by a set of edges. Each node represents a task, and the weight of the node is the execution time of the task. Each edge represents a message transferred from one node to another node, with its weight being the transmission time of the message. Scheduling programs for execution onto processors is a crucial component of a parallel processing system. There are generally two categories of prior art scheduler using DAGs: centralized and decentralized (not shown). An example of a centralized scheduler (10) is shown in FIG. 1 to include a scheduler (30) and a plurality of program execution nodes (12), (14), (16), (18), and (20). The nodes (12), (14), (16), (18), and (20) communicate with each other and the scheduler (30) across a network. In the centralized scheduler (10), an execution request for a program is made to the scheduler (30) which assigns the program to one of the nodes (12), (14), (16), (18) or (20) in accordance with a state of each node. An example of a routine implemented with a centralized scheduler is a first in first out routine (FIFO) in which each program is assigned to a processor in the order in which they are placed in the queue. Problems with FIFO arise when a program in the queue is subject to dependency upon execution of another program. The FIFO routine does not support scheduling a dependent program based upon execution of a prior program. For example, two programs are provided with an execution dependency such that the first program requires a first data input and generates a second data output, and the second program is dependent upon the second data output from the first program execution, and the second program generates a third data output. If the scheduler assigning the programs to one or more processors is running a FIFO routine and the two programs are assigned to execute on two different nodes, the second data output from the first program execution will be on a different node than the second program execution. The second data output will need to be transferred from the node that executed the first program and produce the second data output to the node in which the second program has been assigned for execution. The process of transferring data between nodes consumes resources of both nodes associated with data encryption and decryption. Accordingly, the centralized scheduler results in a decreased utilization of both the first and second processors respectively executing the first and second programs.

In the decentralized scheduler, a plurality of independent schedulers are provided. The benefit associated with the decentralized scheduler is the scalability in a multinode system. However, the negative aspect of the decentralized scheduler is complexity of control and communication among the schedulers to efficient allocate resources in a sequential manner to reduce operation and transmission costs associated with transferring data across nodes for execution of dependent programs. Accordingly, there is an increased communication cost associated with a decentralized scheduler.

There is therefore a need for a method and system to efficiently assign resources based upon a plurality of execution requests for a set of programs having execution dependency with costs associated with data transfer and processing accounted for in a dynamic manner.

SUMMARY OF THE INVENTION

This invention comprises a method and system for dynamically scheduling execution of a program among two or more processor nodes.

In one aspect of the invention a method is provided for assigning resources to a plurality of processing nodes. Priority of execution dependency of a program is decided. In response to the decision, the program is dynamically assigned to a node based upon the priority and in accordance with a state of each node in a multinode system. Preemptive execution of the program is determined, and the program is executed at a designated node non-preemptively in response to a positive determination.

In another aspect of the invention, a system is provided with a plurality of operating nodes, and a scheduling manager to decide priority of execution dependency of a program. A global scheduler is also provided to dynamically assign the program to one of the nodes based upon the priority and a state of each node in the system. In addition, a program manager is provided to determine applicability of preemptive execution of the program, and to non-preemptively execute the program at a designated node in response to a positive determination.

In a further aspect of the invention, an article is provided with a computer-readable signal-bearing medium with a plurality of operating nodes in the medium. Means in the medium are provided for deciding priority of execution dependency of a program. In addition, means in the medium are provided for dynamically assigning the program to one of the nodes based upon the priority and a state of each node in the system. Means in the medium are provided for determining applicability of preemptive execution of the program, and to non-preemptively execute the program at a designated node in response to a positive determination.

Other features and advantages of this invention will become apparent from the following detailed description of the presently preferred embodiment of the invention, taken in conjunction with the accompanying drawings.

BRIEF DESCRIPTION OF THE DRAWINGS

FIG. 1 is a block diagram of a prior art centralized scheduler.

FIG. 2 is a block diagram is a global scheduler according to the preferred embodiment of this invention, and is suggested for printing on the first page of the issued patent.

FIG. 3 is flow chart illustrating a high level operation of processing flow.

FIG. 4 is a flow chart illustrating workflow analysis.

FIG. 5 is a flow chart illustrating assignment of priority to programs in a workflow.

FIG. 6 is a flow chart illustrating logical node assignment.

FIG. 7 is a flow chart illustrating scheduling a program at a node.

FIG. 8 is a flow chart illustrating execution of a program at a node.

DESCRIPTION OF THE PREFERRED EMBODIMENT Technical Details

A grid environment (50) is shown in FIG. 2 and is composed of a global scheduler (60) and a plurality of program execution units (70) and (80), known as nodes. Although only two nodes are shown, more nodes may be addressed to the system. Each node has a program execution unit (72) and (82), respectively, and a local scheduler (74) and (84) that has a local program execution queue (not shown) to manage execution of programs assigned to the respective node. The nodes (70) and (80) communicate with each other and the global scheduler (60) across a local or wide area network (90). An execution request for a program is made to the centralized scheduler (60) which assigns the program to one of the nodes (70, 80) in accordance with a state of each node to execute the program. The centralized scheduler (60) includes a wait queue (62), a workflow database (64), a performance database (66), and an assignment database (68). Each of the nodes (70) and (80) provide processing power, and outputs result of program execution to the global scheduler (60). A web server (not shown) in communication with the global scheduler (60) and each of the nodes (70) and (80) dynamically generates transactions to obtain execution requests and process data. The global scheduler (60) controls processing of a requested program to one or more of the nodes.

FIG. 3 is a flow chart (100) showing a high level processing of program assignments. A workflow submission request, i.e. execution request, is received from a user (11). The workflow is analyzed (112) and executed (114) prior to scheduling an execution of an associated program in the workflow (116). Following execution at step (116) the results are provided to the user (118). There are essentially two procedures to the high level processing. The first procedure is the workflow analysis (112) conducted subsequent to receipt of a workflow submission, and is detailed in FIG. 4. The second procedure involves three components: an execution request for a workflow from the user (114), scheduling and executing programs in the workflow (116), and providing results to the user (118).

As mentioned above, the workflow analysis (112) of FIG. 3 is shown in detail in FIG. 4 (150). The first step of the workflow analysis is assigning priority (152). The program execution priority is decided based on the execution dependency relation of a given program before actual program execution. There are two optional methods for determining priority of assignment of a program. One method is a known as topological sorting, and the second method is based upon the distance from the start program. The topological sorting method involves sorting a directed acyclic graph (DAG) and deciding the priority of the program by incrementing by a factor of 1/(i−1) in sequence, where i indicates the number of programs included in the DAG. The second method, known as the shortest path length, involves computing the distance from the start program, and then deciding the priority as the value normalized by the maximum distance. When there is more than one group of program sets to be executed, the decision on priority of execution is applied to all the program sets to be executed. In either method of assigning priority to a program, the program execution request is added to the global wait queue (62). The entries in the wait queue are sorted based on the priority assigned to the program. When any node is waiting for program execution and the wait queue is not empty, a calculation csts, i.e. cost of assignment to a target node, for program execution is conducted for each program in the queue in order of priority. After execution of the program is completed at the assigned node, an execution request for a subsequent dependent program is added to the wait queue. The entries in the wait queue are rearranged in accordance with the priority assigned to the program. This procedure is repeated until the wait queue is empty.

FIG. 5 is a flow chart (170) illustrating assignment of priority to each program in a group of programs. The first step is a test to determine if there is only one program in the group (172). A positive response to the test at step (172) will result in storing the priority to this one program (174) in the workflow database (64) on the global scheduler (60). However, a negative response to the test at step (172) is an indication that there are at least two or more programs in the group that need to be prioritized. Programs making up a strongly connected component are detected and grouped together (176). The programs grouped in this manner are identified as a strongly connected component group. Other programs that are not part of a strongly connected component are each grouped individually into groups of one program each with the number of programs in each group set as an integer of one. Each of the groups are sequenced by topological sorting (178) with the priority Pi of the i-th group Gi being decided in the following range: 0.0<Pi<1.0, such that Pi−1<P assuming that the priority of a start group is 0.0 and the priority of an end group is 1.0.

Following the sorting process at step (178), priority is assigned to each group (180). The process of assigning priority to each group is applied recursively for each program constituting the strongly connected component group (182) by returning to step (172). The decision of priority Pi is given to group Gi,s and the priority Pi,j is given to the jth group Gi,j in a range of Pi<Pi,j<Pj+1, such that Pi,j<Pj+i in the sequence acquired by topologically sorting the DAG created by excluding the input into Gi,s as the root. The purpose of normalizing the priority of each program is to enable programs in different program sets to be executed with the same presence. That is, when there are nodes for computing and the program sets have an equal total computation time, in situations when program sets request execution at the same time, the computation can be ended at the same time given the equal computation time between the sets. However, in a case where a program set includes a preferential request, the request includes a weight value. The priority assigned to the program is then multiplied by the weight value and applied to the scheduling method described above. Accordingly, it is required that the programs within the groups be recursively split into strongly connected components to decide the priority.

Following the assignment of priority to a group of programs, as well as each program within a group (152), a test is conducted to determine if the program or set of programs can be assigned to a logical node to minimize the transfer of data between programs when analyzing execution dependency (154). The determination at step (154) is based upon whether the computation and/or transmission costs can be estimated. FIG. 6 is a flow chart (200) illustrating the details of the process of assigning one or more programs to a logical node, i.e. a temporary node. Initially workflow data is received (202). Following the receipt at step (202), a cost estimate of program calculation and transmission is estimated (204). From the results of measuring the execution of programs having execution dependency in the past, the relationship between input data size and output data size for the programs in the execution dependency graph and the relationship between input data size and processing costs at a node are estimated. This step focuses on assigning the program with a greater amount of computation to the node of higher performance when the required data transfer overhead is minimal and a plurality of nodes are available. The estimation modeling parameter is made by a regression analysis. The costs can be computed based on program cost assignment, such as data transfer costs and whether the program and required data is cached, and program execution cost, such as the computation amount and the predicted end time. When the estimation at step (240) is complete, the maximum cost, including the computation cost of the dependent programs, is calculated (206). The programs in the workflow are then sorted (208) in the order of the calculated transmission cost. The program(s) are sorted in a hierarchy starting with a program having the highest transmission cost among the programs in consideration (210). If there is a tie between one or more programs having the same transmission cost, the tie is broken based upon the maximum cost of program execution including all dependent programs. Each of the programs are assigned to one or more logical nodes (212) based upon the hierarchical arrangement of the programs from step (210). Accordingly, the logical node assignment is based upon the transmission and/or communication cost of the programs in the queue.

Following the process of calculating the costs associated with execution of a program or group of programs, each of the programs or program groups is assigned to one or more logical nodes (156). The assignment to the logical nodes is stored (158) in the workflow database (64) of the global scheduler (60) and is utilized for scheduling execution of associated programs on actual nodes. FIG. 7 is a flow chart (250) illustrating the process of program scheduling. The first step involves waiting for a next event (252), wherein the event may be a new request arrival event or a node status change event.

Thereafter, the execution condition of the next program is checked and submitted to the queue (254). Step (254) includes providing a priority parameter, i, to a newly executable program. The priority parameter is defined as pi={bi, di, mi}, where bi is the priority given to the entire program, di is the priority based on the dependency relation of each program in the execution dependency, and mi is the priority based on the correspondence relation between the logical node assignment and the actual node assignment. The priority mi has the highest priority when the node to be assigned and the actually assigned node mapped from the logical node for the program are matched. The next highest priority is when the logical node is not assigned to the actual node, and the lowest priority is when the node to be assigned to the program(s) is different from the mapped assignment. The entries in the wait queue are sorted based upon the priority parameters. The sorting is made based upon the following precedence: mi<di, bi, i.e. after the sorting based on mi is complete, the sorting is then based on di, followed by sorting based on bi. Following step (254), a node capable of executing a program or a set of programs is selected (256). The node selection process is based upon prior calculated costs, priority, and availability. A test is then conducted (258) to determine of the node selected at step (256) exists. A negative response to the test at step (258) will result in a return to step (252). However, a positive response to the test at step (258) will result in selection of a program or a set of programs for the transfer from the logical node assignment to the physical node (260). A test is then conducted to determine if the program(s) exist (262). If the response to the test at step (262) is negative, the scheduling returns to step (256). However, if the response to the test at step (262) is positive, a new map is created and the program is assigned to the actual node (264). Thereafter, required data transmission is requested for the program input (266), the program is submitted to the physical node's local queue (268), and a program assignment event is generated (270) followed by a return to step (260). Accordingly, the process of scheduling and executing a program includes mapping the program to an actual node for execution.

FIG. 8 is a flow chart (300) illustrating a process for executing a program after it has been assigned to a physical node for execution. The first step involves waiting for a next event (302), wherein the event may be either a program assignment or data transmission completion. Thereafter, an executable program is selected from the local queue of the physical node (304). A test is conducted to determine if the program exists (306). A negative response to the test at step (306) will return to step (302) for another event. However, a positive response to the test at step (306) will result in executing the selected program within an assigned period (308). A subsequent test is conducted to determine if the program execution has concluded within the assigned time period (310). A negative response to the test at step (310) will return to step (304) to select another program from the queue. However, a positive response to the test at step (310) will remove the executed program from the local queue (312). The performance data generated from the program execution is stored in the performance database (66) of the global scheduler (60). Thereafter, another test is conducted to determine if the destination of the data generated from the program execution has been decided (316). A positive response to the test at step (316) will allow the generated data to be transmitted (318). Thereafter or following a negative response to the test at step (316) a node status change event is generated (320). Following step (320), the process returns to step (304) for selection of a subsequent program from the local queue. Accordingly, the actual node assigned to executed the program stores performance data within the global scheduler.

Advantages Over the Prior Art

The global scheduler dynamically assigns resources while optimizing overhead. Assignment of a workflow to a logical node is employed to mitigate communication and transmission costs associated with execution of a plurality of programs in the workflow by a plurality of nodes in the system. The priority of each program is normalized and sorted in the order of priority. Accordingly, the use of the global scheduler in conjunction with logical node assignments supports cost effective assignment of programs in a workflow to an optimal mode.

Alternative Embodiments

It will be appreciated that, although specific embodiments of the invention have been described herein for purposes of illustration, various modifications may be made without departing from the spirit and scope of the invention. In particular, the assignment of programs in a workflow to a logical node to determine communication and transmission costs may be removed to allow the programs to be forwarded directly to a node having a local queue. Accordingly, the scope of protection of this invention is limited only by the following claims and their equivalents.

Referenced by
Citing PatentFiling datePublication dateApplicantTitle
US7644258Aug 29, 2005Jan 5, 2010Searete, LlcHybrid branch predictor using component predictors each having confidence and override signals
US7831902 *May 31, 2006Nov 9, 2010Microsoft CorporationDisplaying interrelated changes in a grid
US7958058 *Mar 2, 2007Jun 7, 2011International Business Machines CorporationSystem, method, and service for migrating an item within a workflow process
US8028152Oct 31, 2007Sep 27, 2011The Invention Science Fund I, LlcHierarchical multi-threading processor for executing virtual threads in a time-multiplexed fashion
US8266412Oct 31, 2007Sep 11, 2012The Invention Science Fund I, LlcHierarchical store buffer having segmented partitions
US8275976Oct 31, 2007Sep 25, 2012The Invention Science Fund I, LlcHierarchical instruction scheduler facilitating instruction replay
US8296550Oct 31, 2007Oct 23, 2012The Invention Science Fund I, LlcHierarchical register file with operand capture ports
US8539035 *Sep 15, 2009Sep 17, 2013Fujitsu LimitedMessage tying processing method and apparatus
US8621437 *May 1, 2008Dec 31, 2013Intuit Inc.Weighted performance metrics for financial software
US20080133868 *Oct 31, 2007Jun 5, 2008Centaurus Data LlcMethod and apparatus for segmented sequential storage
US20090019259 *Sep 22, 2008Jan 15, 2009Fujitsu LimitedMultiprocessing method and multiprocessor system
US20090216783 *Feb 25, 2008Aug 27, 2009Alexander GebhartHierarchical system operation in an adaptive computing environment
US20090276761 *May 1, 2008Nov 5, 2009Intuit Inc.Weighted performance metrics for financial software
US20110276977 *May 7, 2010Nov 10, 2011Microsoft CorporationDistributed workflow execution
US20120167096 *Mar 6, 2012Jun 28, 2012International Business Machines CorporationManaging the Processing of Processing Requests in a Data Processing System Comprising a Plurality of Processing Environments
US20120266023 *Apr 11, 2012Oct 18, 2012Brown Julian MPrioritization and assignment manager for an integrated testing platform
US20130145187 *Dec 1, 2011Jun 6, 2013Xerox CorporationMulti-device powersaving
US20130254772 *Mar 21, 2012Sep 26, 2013Phillip Morris InternationalVerification of complex workflows through internal assessment or community based assessment
WO2008144239A2 *May 8, 2008Nov 27, 2008Network AutomationAgent workflow system and method
Classifications
U.S. Classification718/100
International ClassificationG06F9/46
Cooperative ClassificationG06F9/5083, G06F9/5038
European ClassificationG06F9/50L, G06F9/50A6E
Legal Events
DateCodeEventDescription
May 25, 2005ASAssignment
Owner name: INTERNATIONAL BUSINESS MACHINES CORPORATION, NEW Y
Free format text: ASSIGNMENT OF ASSIGNORS INTEREST;ASSIGNORS:TANIGUCHI, MASAAKI;KUBO, HARUNOBO;REEL/FRAME:016275/0306;SIGNING DATES FROM 20041105 TO 20041110