Patent Grant
8055940
The present application generally relates to a concurrent computing and more specifically to providing detection of communication error in a concurrent computing environment.
MATLAB® is a product of The MathWorks, Inc. of Natick, Mass., which provides engineers, scientists, mathematicians, and educators across a diverse range of industries with an environment for technical computing applications. As a desktop application, MATLAB® allows scientists and engineers to interactively perform complex analysis and modeling in their familiar workstation environment. With many engineering and scientific problems requiring larger and more complex modeling, computations accordingly become more resource intensive and time-consuming. However, a single workstation can be limiting to the size of the problem that can be solved, because of the relationship of the computing power of the workstation to the computing power necessary to execute computing intensive iterative processing of complex problems in a reasonable amount of time.
For example, a simulation of a large complex aircraft model may take a reasonable amount of time to run with a single workstation with a specified set of parameters. However, the analysis of the problem may also require the model be computed multiple times with a different set of parameters, e.g., at one-hundred different altitude levels and fifty different aircraft weights, to understand the behavior of the model under varied conditions. This would require five-thousand computations of the model to analyze the problem as desired and the single workstation would take an unreasonable or undesirable amount of time to perform these computations. Therefore, it is desirable to perform a computation concurrently using multiple workstations when the computation becomes so large and complex that it cannot be completed in a reasonable amount of time on a single workstation.
In another example, an application can have a mathematical function that is to be integrated in parallel using a quadrature algorithm. In this case, the mathematical function must be evaluated a large number of times in order to calculate the integral to a sufficient degree of accuracy, and each evaluation of the mathematical function may take a large amount of time. To perform the integration in a reasonable amount of time, it would be desirable to have multiple workstations working on the integration in parallel, and communicating partial results with one another until a result with sufficient accuracy is reached.
Applications that are traditionally used as desktop applications, such as MATLAB®, need to be modified to be able to utilize the computing power of concurrent computing, such as parallel computing and distributed computing. Each machine or workstation needs to have its local copy of the application or at least the part of the application that has the necessary functionality for the machine or workstation to perform concurrent computing and the requested computations. Between the different instances of the application, there need to be a way to communicate and pass messages between the machines and workstations so that the multiple machines or workstations in the concurrent computing environment can collaborate with each other.
Message passing is a form of communication used in concurrent computing for different processes on the same or different machines/workstations to communicate with each other in the concurrent computing environment. Communication is made by the sending of messages from one machine/workstation to another machine/workstation. Forms of messages include function invocation, signals, and data packets. One example of a message passing method that establishes a communication channel between machines or workstations is Message Passing Interface (MPI).
When developing concurrent computing programs, such as parallel programs, especially in the “single program, multiple data” model, it is possible to introduce communication mismatches among the multiple nodes in a concurrent computing environment. Communication mismatch can be due to send/receive inconsistency caused by an error in program execution flow, such as a message was not sent because one of the processes exits a loop in an untimely manner. A mismatch can also be due to incorrect sender or receiver. A bug in the parallel program also can cause a communication mismatch. Some errors are non-deterministic, such as ones caused by differences in execution times caused by different data inputs. Errors can easily occur when there is a change in execution environment, such as a change in parallel platform. A communication mismatch in one part of an application may result in errors becoming apparent in a separate part of the application because the communication mismatch may leave some undeliverable messages in a pending state—when these messages are eventually received, they will not be what the receiver expects. A deadlock can possibly occur due to communication mismatch and causes the application to hang. As building many-core multi-processor systems and clusters becomes more popular, debugging a communication mismatch in a concurrent computing program becomes exponentially harder.
Some embodiments of the present invention provide a system and method for detecting communication error among multiple nodes in a concurrent computing environment. A checkpoint called barrier synchronization point is used to check for any communication mismatch in a concurrent computing program. The barrier synchronization point can be placed anywhere in a concurrent computing program as desired. Once a node has reached the barrier synchronization point, the node suspends execution and becomes idle to wait for other nodes to reach the barrier synchronization point as well. No node can leave the barrier synchronization point until all nodes have entered the barrier synchronization point. If a node attempts to initiate communication with another node that has already entered the barrier synchronization point, an error is raised immediately. Once all the nodes have reached the barrier synchronization point, any message in the send/receive buffer is flushed before resuming to normal execution to ensure that any communication mismatch before the barrier synchronization point does not continue past the barrier synchronization point. Each message to be flushed represents a communication mismatch and a warning or an error can be issued.
Alternatively, instead of barrier synchronization points, regions can be used to practice the present invention. Nodes that use the region-based implementation do not suspend execution when they are leaving one region and entering another. In one embodiment of the present invention, each message is packaged with information that identifies the region that the sending node is in so that the receiving node can determine if such message can be successfully received without error. The receiving node checks if the region information in the message is compatible with the region that the receiving node is currently in and an error is raised if there is an incompatibility between the regions. In another embodiment of the present invention, a sending node queries the region that the receiving node is in and compares the region of the receiving node with the region that the sending node is currently in. If the receiving node is in a compatible region with the sending node, then a message is sent from the receiving node to the sending node. However, if the receiving node is in an incompatible region with the sending node, then a communication mismatch is detected. In yet another embodiment of the present invention, a message is sent by a sending node without information on the region that the sending node is in and the receiving node queries the region that the sending node is in before the receiving node successfully receives the message. If the region of the receiving node is compatible with the region of the sending node, then the receiving node successfully receives the message. If the region of the receiving node is incompatible with the region of the sending node, then a communication mismatch is detected.
According to one aspect, a method may include concurrently executing a concurrent computing program by first and second nodes, where the concurrent computing program may include one or more checkpoints; processing, at the first node, an instruction to enter a first one of the one or more checkpoints; suspending execution of the concurrent computing program by the first node when the first node enters the first checkpoint; determining, by the first node, when the second node enters a second one of the one or more checkpoints; and resuming execution of the concurrent computing program by the first node when the second node enters the second checkpoint.
According to another aspect, a node in a system that includes a number of nodes may include means for receiving a concurrent computing program that is to be concurrently executed by the nodes, where the concurrent computing program may include a barrier synchronization point; means for executing the concurrent computing program; means for determining arrival at the barrier synchronization point; means for suspending execution of the concurrent computing program after determining arrival at the barrier synchronization point; and means for resuming execution of the concurrent computing program when at least one other one of the nodes also arrives at the barrier synchronization point.
According to yet another aspect, a computer-readable medium may contain instructions for executing a concurrent computing program that is concurrently executed by a number of nodes. The concurrent computing program may include a number of checkpoints. The computer-readable medium may include instructions for entering one of the checkpoints; instructions for suspending execution of the concurrent computing program after entering the one of the checkpoints; instructions for discarding any undelivered messages after suspending execution of the concurrent computing program; instructions for determining whether another one of the nodes has also entered the one of the checkpoints; and instructions for resuming execution of the concurrent computing program when the other one of the nodes has also entered the one of the checkpoints.
According to a further aspect, a method may include concurrently executing a concurrent computing program by first and second nodes, where the concurrent computing program may include a number of different regions; determining that the first node is currently operating in a first one of the regions of the concurrent computing program; determining that the second node is currently operating in a second one of the regions of the concurrent computing program; and permitting communication between the first and second nodes when the first region is compatible with the second region.
According to another aspect, a system may include a first node and a second node. Each of the first and second nodes may be configured to concurrently execute a concurrent computing program that includes a number of different regions. The first node may be configured to identify one of the regions of the concurrent computing program in which the second node is currently operating, determine whether the identified region is compatible with one of the regions of the concurrent computing program in which the first node is currently operating, and either send a message to the second node or receive a message from the second node when the identified region is compatible with the one of the regions in which the first node is currently operating.
The foregoing and other objects, aspects, features, and advantages of the invention will become more apparent and may be better understood by referring to the following description taken in conjunction with the accompanying drawings, in which:
Some embodiments of the present invention provide a system and method for detecting communication errors among multiple nodes in a concurrent computing environment. A node is used herein to refer to a concurrent computing lab (described later). A barrier synchronization point or regions can be used to check for communication mismatch. The barrier synchronization can be placed anywhere in a concurrent computing program. If a communication error occurred before the barrier synchronization point, it would at least be detected when a node enters the barrier synchronization point. Once a node has reached the barrier synchronization point, it is not allowed to communicate with another node regarding data that is needed to execute the concurrent computing program, even if the other node has not reached the barrier synchronization point. Before leaving the barrier synchronization point and resuming execution of the concurrent computing program, all messages in the send/receive buffer can be flushed so that communication error do not propagate beyond the barrier synchronization point. The concurrent computing nodes can then resume and continue execution of the concurrent computing program and communicate data with each other again. The concurrent computing application does not need to be aborted when a communication error is detected. In another aspect of the present invention, a program may be divided into regions instead of using the barrier synchronization points to detect a communication mismatch. If a node attempts to communicate with another node that is currently in an incompatible region, then an error is raised to signal the communication mismatch.
The following illustrative embodiments will be described for illustrative purposes relative to a MATLAB®-based technical computing environment. Although the illustrative embodiment will be described relative to a MATLAB®-based application, one of ordinary skill in the art will appreciate that the present invention may be applied to parallel or distributed processing of technical or non-technical computing tasks with other technical or non-technical computing environments. Examples of technical computing environments that are suitable to practice with the present invention include ones from software products of LabVIEW® or MATRIXx from National Instruments, Inc., or Mathematica® from Wolfram Research, Inc., or Mathcad of Mathsoft Engineering & Education Inc., or Maple™ from Maplesoft, a division of Waterloo Maple Inc, or R from The R Project for Statistical Computing of Australia, or Scilab from Institut national de recherche en informatique et en automatique (INRIA) of France, or Octave from University of Wisconsin of Madison, Wis. or products for embedded systems from OPAL-RT Technologies, Inc. of Ann Arbor, Mich.
Some embodiments of the present invention can also be applied to an embedded control system or a network system of embedded controls. An embedded control system includes a physical system that is being controlled by computing hardware, such as a microprocessor. The computing hardware is embedded in the physical system by means of sensors and actuators. One embedded control system may contain multiple computing hardware entities and these hardware entities communicate to achieve a desired behavior of the physical system. These hardware entities can work in a concurrent computing environment (discussed below). The communication among the multiple computing hardware entities can synchronize in time and the communication of information can be done periodically or aperiodically. For the overall system to behave as desired, the communication needs to be validated and missed points of communication need to be identified and the incorrectly communicated information needs to be acted upon, such as discarding the incorrectly communicated information and issuing a warning.
Some embodiments of the present invention may apply to concurrent processing of streaming data. The data can be processed in real-time. Alternatively, data can be stored first before it is being processed. For example, a fast Fourier transform may be used to analyze the frequency spectrum of the measured values from a sensor. In case of the on-line processing, real-time processing is typically performed. Real-time processing may be employed in off-line processing. The measured values from a sensor may be supplied to the nodes in the concurrent computing environment as a stream of data and the processing may be performed by a number of concurrent computations that exploit the communication error and warning detection of the present invention.
The memory 106 may comprise a computer system memory or random access memory such as DRAM, SRAM, EDO RAM, etc. The memory 106 may comprise other types of memory as well, or combinations thereof A user may interact with the computing device 102 through a visual display device 114 such as a computer monitor, which may include a user interface 115. The computing device 102 may include other I/O devices such a keyboard 110, a touchscreen, a camera, a microphone, and a pointing device 112, for example a mouse, for receiving input from a user. Optionally, the keyboard 110 and the pointing device 112 may be connected to the visual display device 114. The computing device 102 may include other suitable conventional I/O peripherals. The computing device 102 may further comprise a storage device 108, such as a hard-drive, CD-ROM, or other computer readable media, for storing an operating system 116 and other related software, and for storing application 120, such as parallel computing with MATLAB® or distributed computing with MATLAB®.
Application 120 may be running a concurrent computing environment 122 to enable concurrent computing on the computing device 102. The application 120 can also include a communication error detector 128 for detecting a communication error among multiple nodes in the concurrent computing environment 122. When a communication error is detected, a warning can be issued. One of ordinary skill in the art will appreciate that communication error detector 128 can be adapted to be included as part of the application 120, or it can be a stand-alone application/module/script/program that responds to calls from the application 120, such as communication error detector 128′. Additionally, the operating system 116 and application 120 can be run from a computer readable media, such as, for example, KNOPPIX®, a bootable CD for GNU/Linux.
The computing device 102 may include a network interface 118 to interface to a Local Area Network (LAN), Wide Area Network (WAN) or the Internet through a variety of connections including, but not limited to, standard telephone lines, LAN or WAN links (e.g., 802.11, T1, T3, 56 kb, X.25), broadband connections (e.g., ISDN, Frame Relay, ATM), wireless connections, controller area network (CAN), or some combination of any or all of the above. The network interface 118 may comprise a built-in network adapter, network interface card, PCMCIA network card, card bus network adapter, wireless network adapter, USB network adapter, modem or any other device suitable for interfacing the computing device 102 to any type of network capable of communication and performing the operations described herein. Moreover, the computing device 102 may be any computer system such as a workstation, desktop computer, server, laptop, handheld computer or other form of computing or telecommunications device that is capable of communication and that has sufficient processor power and memory capacity to perform the operations described herein.
The computing device 102 can be running any operating system such as any of the versions of the Microsoft® Windows® operating systems, the different releases of the Unix and Linux operating systems, any version of the MacOS® for Macintosh computers, any embedded operating system, any real-time operating system, any open source operating system, any proprietary operating system, any operating systems for mobile computing devices, or any other operating system capable of running on the computing device and performing the operations described herein. The operating system may be running in native mode or emulated mode.
The concurrent computing client 250 can be a technical or non-technical computing software application. Concurrent computing client 150 may provide a technical computing and/or graphical modeling environment for generating block diagram models and to define mathematical algorithms for simulating models. The concurrent computing client 250 may include all or a portion of the functionality provided by the stand-alone desktop application of MATLAB®. Additionally, the concurrent computing client 250 can be any of the software programs available in the MATLAB® product family. Furthermore, the concurrent computing client 250 can be a custom software program or other software that accesses functionalities of software programs in the MATLAB® product family via an interface, such as an application programming interface (API), or by other means. One of ordinarily skill in the art will appreciate the various combinations of client types that may access the functionalities of the system.
In one embodiment of the present invention, concurrent computing client 250 is also a concurrent computing lab. In such a configuration, communication channels are set up among the concurrent computing labs (concurrent computing client 250 and concurrent computing labs 270A-N). Each of the concurrent computing labs (including the concurrent computing client 250) has its local copy of a computer program that is executed in the corresponding concurrent computing labs, so there is no main concurrent computing lab that distributes executions to the other concurrent computing labs. Alternatively, a copy of the computer program can be accessed through a network connection. The local copy of the program for each lab may or may not be identical. The concurrent computing client 250 can additionally have the functionality to accept inputs and/or commands from a user related to the computer program using a tool such as an Integrated Development Environment (IDE). The concurrent computing client 250 and concurrent computing labs 270A-N can be configured to perform distributed computing or parallel computing.
In one embodiment of the present invention, functions can be defined, by the concurrent computing client 250 with an application programming interface (API) and/or programming language, representing a technical computing task to be executed by either a technical computing environment local to the client 150, or remote on the workstations 270A-N. Tasks can be declared on a concurrent computing client 250 and additionally organized into jobs. A job is a logical unit of activities, or tasks that are processed and/or managed collectively. A task defines a technical computing command, such as a MATLAB® command, to be executed, and the number of arguments and any input data to the arguments. A job is a group of one or more tasks.
In one aspect of the present invention, a task can be directly distributed by the concurrent computing client 250 to one or more computing resources, such as workstations 170A-N. A computing resource performs technical computing on a task and may return a result to the concurrent computing client 250.
In another aspect of the present invention, the system includes a server 160 on which a scheduler 260 runs. The scheduler 260 can be a scheduler provided with application 120, a generic scheduler, or a third-party scheduler that is designed and provided by a company or individual that may not provide application 120. For example, given that application 120 is parallel computing with MATLAB® by The MathWorks, Inc. of Natick, Mass., a third-party scheduler can be MPI Exec, LSF, Condor, Microsoft Compute Cluster Server, or PBS. The server 160 communicates over a network communication channel 130 on the network 140 to the workstations 170A-N. One of ordinary skill in the art will appreciate that any of the workstations 170A-N may include more than one technical computing lab to practice the present invention. Additionally, client 150 and server 160 may also include one or more concurrent computing labs.
The scheduler 260 comprises one or more application software components to provide for the automatic distribution of tasks from the concurrent computing client 250 to one or more of the concurrent computing labs 270A-N. The scheduler 260 allows the concurrent computing client 250 to delegate the management of task distribution to the scheduler 260. The scheduler may also set up for concurrent computing client 250 the concurrent computing labs 270A-N by using the information received from the concurrent computing client 250 regarding the number of concurrent computing labs needed and other configuration information. Hence, the concurrent computing client 250 does not need to know the specifics of the concurrent computing labs 270A-N. The concurrent computing client 250 can define a function to submit the task to the scheduler 260, and get a result of the execution of the task. As such, the scheduler 260 provides a level of indirection between the concurrent computing client 250 and the concurrent computing labs 270A-N.
This eases the distributed programming and integration burden on the concurrent computing client 250. The concurrent computing client 250 does not need to have prior knowledge of the availability of the workstations 170A-N. For multiple task submissions from the concurrent computing client 250, the scheduler 260 can manage and handle the delegations of the tasks to the concurrent computing labs 270A-N and hold the results of the tasks on behalf of the concurrent computing client 250 for retrieval after the completion of technical computing of all the tasks distributed by concurrent computing client 250 or at desired intermediate points. In an alternative implementation, the concurrent computing labs 270A-N may provide concurrent computing client 250 directly the results of the tasks assigned to concurrent computing labs 270A-N by the scheduler 260. The scheduler 260 can further include an object-oriented interface to provide control of delegating tasks and obtaining results in the system 200. The scheduler 260 also provides an interface for managing a group of tasks collectively as a single unit called a job, and on behalf of a concurrent computing client 250, submitting those tasks making up the job, and obtaining the results of each of the tasks until the job is completed. One of ordinary skill in the art will recognize that the functions and operations of the scheduler 260 can be separated into various software components, applications and interfaces. Additionally, the functions and operations of the scheduler 260 may reside on either the concurrent computing client 250 or one of the concurrent computing labs 270A-N instead of the server 160.
Additionally, each of the client 150, the server 160, and the workstations 170A-N can be running the same or different operating systems with the same or different processors. For example, the client 150 can be running Microsoft® Windows®; the server 160 can be running a version of Unix, and the workstations 170A-N a version of Linux. Alternatively, each of the client 150, the server 160 and the workstations 170A-N can be running Microsoft® Windows®. One of ordinarily skill in the art will recognize the various combinations of operating systems and processors that can be running on any of the computing devices (client 150, server 160, workstations 170A-N). One or ordinary skill in the art will also appreciate that some computing device may not have an operating system. For example, an FPGA without an operating system can be configured to perform computations synchronously or asynchronously and put the data on a communication bus.
In one embodiment of the present invention, the collaboration is dynamic. In other words, a user can modify or change the size of the collaboration by adding another computing resource. The user may be provided on the client 150 with a user interface to modify or change the size of the collaboration or designate a specific resource to add or remove from the collaboration. In another embodiment of the present invention, the client 150 can forward the information to the scheduler 260, which will determine a concurrent computing lab to be added or removed from the collaboration.
In step 404, the first node obtains an instruction to enter a barrier synchronization point. The barrier synchronization point can be inserted manually by a user into the concurrent computing program. Alternatively, the barrier synchronization point can be inserted automatically by the communication error detector 128. In one embodiment of the present invention, the positions to insert the barrier synchronization points can be dynamically defined. A user can use a rule-based system to define where barrier synchronization points are inserted. The barrier synchronization point is used to detect if there is any communication error among the nodes. If a communication error is detected, a warning can be issued. The warning can be accompanied by an error code that helps a user to identify the type of communication error that was detected. When a node reaches the barrier synchronization point, the node waits for other nodes in the concurrent computing environment to reach the barrier synchronization point before continuing the execution of the concurrent computing program so that the communication among the nodes can be synchronized. For example, the following pseudo code shows how a barrier synchronization point can be used in one embodiment of the present invention.
In one embodiment of the present invention, the barrier synchronization point is different for each node. In other words, the nodes can execute a different number of lines of code before reaching a synchronization point. Hence, synchronization points can be assigned differently to different nodes to ensure the correct data are being communicated among the nodes to avoid data mismatch. In another embodiment of the present invention, a node only needs to wait for the nodes that the node communicates with when the node reaches a synchronization point. There can be multiple barrier synchronization points in a single concurrent computing program. In one embodiment of the present invention, a barrier synchronization point is represented using a single instruction, such as a single function call. Alternatively, the barrier synchronization point can be represented using a series of instructions. One of ordinary skill in the art will recognize that every barrier synchronization point does not need to be the same. In other words, every barrier synchronization point can be implemented differently.
In step 406, the first node suspends execution of the concurrent computing program. The first node then optionally sends a message to the other nodes in the concurrent computing environment that it has entered the barrier synchronization point in step 408. For example, the following function can be used by a node to inform other nodes in the concurrent computing environment that it has reached the barrier synchronization point.
The first node then checks if all the other nodes have entered the barrier synchronization point in step 410. If all the other nodes have entered the barrier synchronization point, then the first node resumes execution of the concurrent computing program in step 418. If all the other nodes have not entered the barrier synchronization point, then the first node waits for all the other nodes to enter the barrier synchronization point before resuming execution of the concurrent computing program in step 412. For example, the following function can be used by the first node to wait for other nodes to reach the barrier synchronization point.
Once all the nodes have entered the barrier synchronization point, any undelivered messages may be flushed in step 414 to ensure that before resuming to normal execution of the concurrent computing program in step 418, there are no messages buffered by the communication layer. Each message to be flushed represents a communication mismatch and a warning is issued. Moreover, any message received while the first node is waiting at the barrier synchronization point is discarded in step 416 before the first node resumes execution of the concurrent computing program in step 418. For example, the following function can be used to clean up any messages that are sent/received while waiting at the synchronization point.
Hence if there is an error due to communication mismatch, it would at least be detected when a node enters the barrier synchronization point. Before leaving the barrier synchronization point and resuming execution of the concurrent computing program, all messages in the send/receive buffer are flushed so that communication errors do not propagate beyond the barrier synchronization point. The concurrent computing nodes can then resume and continue execution of the concurrent computing program and communicate data with each other again. The concurrent computing application does not need to be aborted when a communication error is detected.
In one embodiment of the present invention, a state of the first node at the synchronization point is stored in step 420. When a communication mismatch is detected after the synchronization point, then the first node may return to the stored state. There may be multiple synchronization points used in a single concurrent computing program, then when a communication mismatch is detected, the concurrent computing nodes can return to the last synchronization point and re-execute the portion of the program where a communication mismatch is detected. Alternatively, the concurrent computing nodes can return the last synchronization point and allow a user to fix any error in the concurrent computing program before re-executing the portion where a communication mismatch is detected.
The following is an example of how a node can receive a message from another node using one embodiment of the present invention.
If the first node has already reached the barrier synchronization point before it can receive or send data, then the second node would detect a communication error in step 512. Optionally, the second node can receive a message from the first node informing that the first node has entered the barrier synchronization point in step 510. This can occur before or after the second node attempts to communicate with the first node. If the first node that the second node wants to communicate with has already reached a synchronization point, then the second node would eventually obtain an instruction to enter the barrier synchronization point in step 514 as well. Then the second node would suspend execution of the concurrent program in step 516 until all nodes have reached the barrier synchronization point.
In another embodiment of the present invention, regions can be defined within a concurrent computing program to detect a communication mismatch. In one preferred embodiment of the present invention, the nodes do not suspend execution of their programs when they exit one region and enter another. The local copy of the concurrent program on each node is separated into one or more regions. One of ordinary skill in the art will appreciate that the overall program being executed can be the same on each node, but the exact code that is executed on each node may be different, such as conditional code that may require one lab to execute one portion of the program while another lab executes another portion. Alternatively, each node may execute a different program providing that the region definitions between the different programs are compatible. The local copies of the same program on different nodes may or may not have the same regions definition, but the regions definition on each node must be compatible. For example, given a program with a conditional code that has 10 lines of code, a first node may assign the first five lines to a first region while a second node may assign the next three lines of code to the same first region.
Two nodes can only communicate with each other when their regions are compatible. If their regions are not compatible, then there is a communication mismatch. In one embodiment of the present invention, each message communication that is sent would include information on the region the sending node is in. When the receiving node receives the message, it checks if the message has region information compatible with the current region of the receiving node. If the receiving node is in an incompatible region compared to the region information in the message, then an error occurs. Each communication is associated with a particular region and only nodes with compatible regions can communicate with each other.
In one embodiment of the present invention, when a node changes from a first region to a second region, a message is sent to other nodes in the concurrent computing environment to inform the new region that the node is in. In this case, if a receiving node with the current region as the first region is waiting to receive a message from the node that changes from the first region to the second region, then the receiving node knows that there is an error since the node has finished execution in the first region and the receiving node will never receive a message from the node while the node is in the first region. However, it is possible for a receiving node to start waiting to receive a message from a sending node when the sending node is still in an incompatible preceding region. At this point, an error is not raised because sending node can possibly get to the right region and send a message to the receiving node. By the time the receiving node actually receives a message from the sending node, the message can possibly be sent while the sending node is in a compatible region with the receiving node. If the region information of the sending node is packaged with the message, then the receiving node can check if the region information in the message is compatible with the current region that the receiving node is in. If the regions are not compatible, then an error is raised at this point. In another embodiment of the present invention, each node can optionally keep a record of the current regions that the other nodes are in. When regions are used to detect a communication mismatch, the nodes do not need to suspend execution when it leaves one region and enters another.
In one embodiment of the present invention, a compatible region means the same region. Alternatively, the definition of a compatible region can be different for each region in the concurrent computing program or concurrent computing node. One of ordinary skill in the art will appreciate that there can be many different ways in defining what a compatible region is and the specific examples mentioned herein should not be used to limit the scope of the present invention.
In one embodiment of the present invention, regions are automatically defined by the communication error detector 128 or application 120. The regions may also be manually defined by a user in the concurrent computing program. A counter variable or comments can be used to define regions. A function call can also be used to define the boundaries between regions, such as labRegionBoundary in the following example. A ruled-base system can also be used to define the regions. One of ordinary skill in the art will appreciate that there are many ways that one can define regions and the specific examples listed herein should not be used to limit the scope of the present invention.
In one embodiment of the present invention, the regions can have a temporal aspect. A region may be defined relative to time. In such a case, nodes can only communicate with each other, for example using MPI/RT, if they are in a compatible temporal region. The definition of temporal region can be different on each node in the concurrent computing environment as long as the definitions of temporal regions are compatible. One of ordinary skill in the art will appreciate that a barrier synchronization point can also be defined relative to time. For example, the end of a 10 millisecond fragment of data can be used as a barrier synchronization point. If one of the labs returns a result before the 10 milliseconds have passed, this may indicate that an error has occurred. One of ordinary skill in the art will appreciate that other operations may be performed on data with a real-time aspect.
In one embodiment of the present invention, each region has an identification number. The identification number can be an integer or other suitable numerical representation. Alternatively, each region can also be identified using alphabets, characters, symbols, or other suitable representation. One of ordinary skill in the art will appreciate that the scope of the present invention is not limited to the specific identification system for the regions mentioned herein and other suitable representation can also be used to identify the different regions.
The following pseudo code illustrates one way of implementing regions using the present invention. The following example assumes that MPI is used with MATLAB® to implement regions.
Many alterations and modifications may be made by those having ordinary skill in the art without departing from the spirit and scope of the invention. Therefore, it must be expressly understood that the illustrated embodiments have been shown only for the purposes of example and should not be taken as limiting the invention, which is defined by the following claims. These claims are to be read as including what they set forth literally and also those equivalent elements which are insubstantially different, even though not identical in other respects to what is shown and described in the above illustrations.
This application is a continuation of U.S. patent application Ser. No. 11/488,432, filed Jul. 17, 2006, the entire content of which is incorporated by reference herein.
| Number | Name | Date | Kind |
|---|---|---|---|
| 4816989 | Finn et al. | Mar 1989 | A |
| 4914657 | Walter et al. | Apr 1990 | A |
| 5768538 | Badovinatz et al. | Jun 1998 | A |
| 5987477 | Schmuck et al. | Nov 1999 | A |
| 6029205 | Alferness et al. | Feb 2000 | A |
| 6216174 | Scott et al. | Apr 2001 | B1 |
| 6336161 | Watts | Jan 2002 | B1 |
| 6430600 | Yokote | Aug 2002 | B1 |
| 6651242 | Hebbagodi et al. | Nov 2003 | B1 |
| 6718484 | Kodera | Apr 2004 | B1 |
| 6834358 | Korenevsky et al. | Dec 2004 | B2 |
| 7117248 | Jordan, Jr. | Oct 2006 | B1 |
| 7191294 | Nakamura et al. | Mar 2007 | B2 |
| 7305582 | Moser et al. | Dec 2007 | B1 |
| 7610510 | Agarwal et al. | Oct 2009 | B2 |
| 20050050374 | Nakamura et al. | Mar 2005 | A1 |
| 20050278620 | Baldwin et al. | Dec 2005 | A1 |
| 20060200730 | Daugherty | Sep 2006 | A1 |
| 20070150714 | Karstens | Jun 2007 | A1 |
| 20070174484 | Lussier et al. | Jul 2007 | A1 |
| 20070260909 | Archer et al. | Nov 2007 | A1 |
| Number | Date | Country | |
|---|---|---|---|
| 20090006621 A1 | Jan 2009 | US |
| Number | Date | Country | |
|---|---|---|---|
| Parent | 11488432 | Jul 2006 | US |
| Child | 11879383 | US |
