KR20130080721A - Host node and memory management method for cluster system based on parallel computing framework - Google Patents

Abstract

PURPOSE: Host node and memory management method are provided to reduce a bug incidence rate of an application by executing an application of a parallel computing programming model between a host node and a calculation node in cluster environment. CONSTITUTION: A host thread (221) executes a host program for a parallel computing framework. The host thread inserts a command related to a kernel program for the parallel computing framework to a command queue. A command scheduler (222) schedules the command queue and selects a command to be executed. The command scheduler generates a request message requesting execution of a command. A memory management unit (280) judges whether having a buffer objet using a command selected by a first calculation module or not. The memory management unit copies a buffer object of a second calculation module different from the first calculation module into the first calculation module. [Reference numerals] (220) Host node; (241a) Calculation module #1; (241b) Calculation module #2; (241c) Calculation module #3; (241d) Calculation module #4; (AA,240) Calculation node; (BB,CC,DD,EE) Calculation module

Description

Host node and memory management method for cluster system based on parallel computing framework}

Related to the Parallel Computing Framework technology.

Parallel computing is a method of operation that performs many computations at the same time. Parallel computing is often used to solve large and complex problems in parallel and solve them in parallel. There are many ways and kinds of parallel computing. Examples include bit level, instruction level, data, and task parallelism. Parallel computing has long been used primarily for high-performance computing, and has come to the fore since the problem consciousness has increased as the physical limits of processor frequencies have been approached. In recent years, with the growing interest in heat generation and power consumption in computer use, it has attracted attention as a powerful paradigm in computer architecture with a multi-core processor as the core.

The representative framework for parallel computing is OpenCL. OpenCL is a framework that allows you to write programs with data and task parallelism that must run on multiple platforms with multiple CPUs, GPUs, and other processors. OpenCL includes a new language based on C99. With this language you can write kernels, APIs, and so on that define and control hybrid platforms. OpenCL provides mission-based parallel or data-based parallel computing.

Typically OpenCL is for single node systems. Recently, however, a cluster system having a large number of nodes has attracted attention, and OpenCL does not work in such a cluster environment. In other words, in order to write an application for OpenCL in a cluster environment, it is inconvenient for a user to use a message programming library for a network. This makes writing an application difficult. Furthermore, OpenCL applications with user-added code may not work on a single node system.

A host node of a cluster system in which the parallel computing framework can operate effectively and a memory management method for such a system are presented.

A host node according to an aspect of the present invention is a host thread for executing a host program for a parallel computing framework, and inserting a command related to a kernel program for the parallel computing framework into a command queue, a command to be executed by scheduling a command queue. A command scheduler for generating a request message requesting execution of the selected command, transmitting a generated request message over the network to a calculation node including a calculation module to execute the command, and a calculation module for executing the selected command. 1. A memory for determining whether a calculation module has a buffer object used by a selected command, and according to the determination result, a memory for copying a buffer object of a second calculation module, which is a different calculation module from the first calculation module, to the first calculation module. It may include a management unit.

In addition, the memory management method according to an aspect of the present invention, executing a host program for the parallel computing framework, inserting a command related to the kernel program for the parallel computing framework to the command queue, scheduling the command queue to be executed Selecting a command, determining whether the first calculation module, which is the calculation module to execute the selected command, has a buffer object used by the selected command, and according to the determination result, a second calculation that is a calculation module different from the first calculation module; Copying the buffer object of the module to the first calculation module, generating a request message requesting execution of the selected command, and transmitting the generated request message to the calculation node including the first calculation module via a network. It may include.

Since the host node and the compute node in the cluster environment send and receive messages to execute an application of the parallel computing programming model, the inconvenience of having to add code such as a message passing interface (MPI) directly to the application is eliminated. This reduces application bugs and increases programmer productivity and portability of parallel computing frameworks.

In addition, in a cluster environment based on a parallel computing framework, a calculation module list is used to determine in advance whether a calculation module that executes a command has a buffer object used by the command, and then buffer objects of another calculation module are transferred to the calculation module. Since it is copied, efficient memory management can be guaranteed.

1 illustrates a configuration of a cluster system according to an embodiment of the present invention.
2 illustrates a configuration of a host node and a compute node of a cluster system according to an embodiment of the present invention.
3 illustrates a memory management method of a cluster system according to an embodiment of the present invention.

Hereinafter, specific examples for carrying out the present invention will be described in detail with reference to the accompanying drawings.

1 illustrates a configuration of a cluster system according to an embodiment of the present invention.

Referring to FIG. 1, the cluster system 100 according to the present embodiment includes a host node 101 and a plurality of computing nodes 102a, 102b, 102n. The host node 101 and each computing node 102a, 102b, 102n are connected via a network 103. Network 103 may be a wired or wireless network. The host node 101 and the computing nodes 102a, 102b, and 102n can exchange information through network messages. The network message may have a form such as a LAN, IPv4, or IPv6 packet.

Cluster system 100 is formed based on a parallel computing framework. In other words, the parallel computing framework according to the present embodiment may operate in the cluster system 100. As the parallel computing framework of the cluster system 100, OpenCL, OpenMP, CUDA, and the like may be used.

An application for the parallel computing framework of the cluster system 100 may be composed of a host program and a kernel program. The host program is a part that manages the kernel program so that the kernel program is executed, and the kernel program may be a part where data operations are processed. According to one aspect, the host program may be executed at the host node 101 and the kernel program may be executed at the compute nodes 102a, 102b, 102n.

The host node 101 executes a host program. The host node 101 may manage the execution of a kernel program by executing a host program of an application created through a parallel computing programming tool. For example, the host node 101 generates a request message for requesting command execution using a command related to a kernel program, and generates the generated request message to the computing nodes 102a, 102b, and 102n via the network 103. It is possible to transmit.

Computing nodes 102a, 102b, 102n execute kernel programs. Computing node (eg, 102a) that receives a request message for requesting command execution from host node 101 executes a command according to the request message, generates a completion message indicating completion of command execution, and generates the generated completion message. It is possible to transmit to the host node 101 via the network 103.

2 illustrates a configuration of a host node and a compute node of a cluster system according to an embodiment of the present invention.

Referring to FIG. 2, the host node 220 may include a host thread 221, a command scheduler 222, command-queues 223a-223f, an issue-queue 224, and a completion-queue 225. Can be.

The host thread 221 executes a host program.

The host thread 221 inserts a command into the command queues 223a to 223f. Each command-queue 223a-223f may correspond to each calculation module 241a, 241b, 241c, 241d of the calculation nodes 240, 240b. For example, the first command-queue 223a in the first command-queue group 226 may be mapped to the calculation module # 1 241a in the first compute node 240. The command may be various operations related to the kernel program to be executed in the calculation node 240. In other words, the host thread 221 determines the calculation node 240 and the calculation module 241a to execute a command related to the kernel program, and the command-queue corresponding to the determined calculation node 240 and the calculation module 241a is determined. It is possible to put a command in 223a).

Command scheduler 222 schedules command-queues 223a through 223f. The round-robin method may be used as the schedule policy, but various schedule policies may be applied depending on the application purpose. For example, the command scheduler 222 may select a command to be executed according to a predetermined schedule policy, and retrieve the selected command from the command queues 223a to 223f.

The command scheduler 222 generates a request message using the selected command. The request message is a packet type message transmitted through a network, and may be a message for requesting execution of a selected command. For example, the command scheduler 222 may generate packet data having an address field and a description field of the command from the calculation node 240 and the calculation module 241a mapped to the command-queue 223a of the selected command. have.

The command scheduler 222 sends the generated request message to the computing node 240 via the network 260.

The memory manager 280 may use a buffer object used by a command selected by a calculation module (hereinafter, referred to as an execution calculation module) (for example, 241a) to execute a command selected by the command scheduler 222 (hereinafter referred to as a 'necessary buffer object'). Determine whether or not it has

A buffer object may be defined as an abstracted memory area used by an application for a parallel computing framework. This buffer object can be mapped to multiple memories without being dependent on a particular memory. Therefore, it is possible for several calculation modules 241a, 241b, 241c, and 241d to share one buffer object. A required buffer object may be a buffer object required for an execution calculation module to execute a command.

The memory manager 280 may determine whether the execution calculation module 241a holds the required buffer object by using a predetermined calculation module list. The calculation module list is a list of calculation modules configured to hold the latest data about the buffer object for each buffer object. The memory manager 280 may determine whether the execution calculation module holds the required buffer object according to whether the execution calculation module 241a exists in the calculation module list of the required buffer object.

As a result of determination, when the execution calculation module 241a does not hold the required buffer object, the memory manager 280 searches for another calculation module (eg, 241b) that holds the required buffer object, and retrieves the found calculation module 241b. The necessary buffer object can be copied from the memory module and written to the memory of the execution calculation module 241a. The other calculation module 241b holding the required buffer object may be any one of the calculation modules present in the calculation module list of the required buffer object.

According to one aspect, if there are a plurality of calculation modules in the calculation device list, it is possible to select a calculation module to copy the data in the following priority order.

1. CPU on the same node

2. GPU on same node

3. CPU on another node

4. GPUs on other nodes

This is because data movement on the network can only be performed between nodes and main memory. Therefore, if you want to move memory from one node's GPU to another node's GPU, copy the data from the GPU memory to move to the node's main memory and then copy the data to the other node's main memory over the network. After that, the data must be copied from the node's main memory to the GPU memory. It is desirable to prioritize the CPU computing device to reduce this additional copying cost.

In addition, the memory manager 280 may add the execution calculation module to the calculation module list when the required buffer object is copied to the execution calculation module. According to a further aspect, when the buffer object is modified, the memory manager 280 then clears the calculation module list of the buffer object, and then calculates the calculation module (eg, 241a) that modified the buffer object. You can also add The memory manager 280 may analyze the kernel source (eg, the OpenCL kernel source) to find out which command (or which kernel) modifies which buffer object.

In FIG. 2, assume that the command selected by the scheduler 222 is a command in the first command queue 223a, and the calculation module to execute this command is calculation module # 1 241a. Also assume that calculation module # 2 241b and calculation module # 3 241c exist in the calculation module list for the buffer object 'A' used by the selected command. The memory manager 280 determines whether the calculation module # 1 241a is in the calculation module list related to the buffer object A. FIG. If the calculation module # 1 241a is in the calculation module list for the buffer object A, the command scheduler 222 sends a request message to the calculation module # 1 241a. If there is no calculation module # 1 241a in the calculation module list for the buffer object A, the memory management unit 280 corresponds to the calculation module # 2 (241b) or calculation module # 3 (241c) in the calculation module list. Copy the buffer object to compute module # 1 241a. The command scheduler 222 then sends a request message to the calculation module # 1 241a.

According to an aspect, the command scheduler 222 may generate a request message after the required buffer object is copied to the execution calculation module by the memory manager 280 when the required buffer object does not exist in the execution calculation module. .

According to another aspect, the command scheduler 222 may store a command corresponding to the transmitted request message in the issue-queue 224 and wait for a response of the calculation node 240. If the compute node 240 responds to completion of the command execution, the command scheduler 222 may remove the command stored in the issue-queue 224 and move it to the completion-queue 225.

In addition, in FIG. 2, the calculation node 240 includes a number of calculation modules 241a, 241b, 241c, 241d, a command handler 242, a number of ready-queues 243a, 243b, 243c, 243d, and a number of modules. Threads 244a, 244b, 244c, 244d.

Each calculation module 241a, 241b, 241c, and 241d may be a device for computing and processing data, such as a CPU and a GPU.

The command handler 242 receives a request message from the host node 220 via the network 260.

The command handler 242 generates a command object according to the received request message, and puts the generated command object into the preparation queues 243a, 243b, 243c, and 243d. The preparation-cues 243a, 243b, 243c, and 243d correspond to the respective calculation modules 241a, 241b, 241c, and 241d. Therefore, if it is determined that the command is executed in the first calculation module 241a according to the request message, the command object may be inserted into the preparation-queue 243a of the first calculation module 241a.

The module threads 244a, 244b, 244c and 244d execute commands using the command objects inserted into the preparation-queues 243a, 243b, 243c and 243d and the calculation modules 241a, 241b, 241c and 241d. For example, in the above example, it is possible for the module thread 244a of the first calculation module 241a to withdraw the command object from the ready-queue 243a so that the command is processed in the first calculation module 241a based on it. .

According to one aspect, the module threads 244a, 244b, 244c, 244d may insert the completed command into the completion-queue 245 when the command execution is complete.

According to a further aspect, the command handler 242 uses a command in the completion-queue 245 to generate a completion message indicating the completion of the command execution, and the generated completion message via the network 260 to the host node 220. It is possible to send

Each configuration illustrated in FIG. 2 may be implemented in the form of electrical circuits and / or hardware, or may be implemented in the form of an application program executed in a predetermined processor. In addition, the division of each structure is only an example of the logical division according to the function. Therefore, the functions may be distinguished according to different criteria from those shown in FIG. 2. In other words, two or more functional units may be integrated into one functional unit, or some of the functions performed in one functional unit may be performed in one or more other functional units.

3 illustrates a memory management method of a parallel computing framework based cluster system according to an embodiment of the present invention. This can be done by the host node of FIG.

2 and 3, according to the memory management method according to the present embodiment, a host program for the parallel computing framework is first executed, and commands related to the kernel program for the parallel computing framework are inserted into the command queue ( 301). For example, it is possible for the host thread 221 to put a command in the command queue 223a.

The command to be executed is selected according to the scheduling of the command queue (302). For example, it is possible for the command scheduler 222 to select a command in the command queue 223a in accordance with a predetermined schedule policy.

In operation 303, it is determined whether the calculation module to execute the selected command has a buffer object used by the command. For example, the memory manager 280 may determine whether the calculation module exists in the calculation module list of the buffer object by using the calculation module list for each buffer object. If the calculation module (eg, 241a) that executes the selected command does not have a corresponding buffer object, the buffer object of another calculation module (eg, 241b) present in the calculation module list is used. Is added to the memory, and the calculation module 241a is added to the calculation module list to execute the selected command (304).

A request message requesting execution of the selected command is transmitted to the corresponding calculation node (305). For example, the command scheduler 222 may generate a request message for requesting execution of the selected command and transmit the generated request message through the network.

As described above, according to the disclosed embodiments, in a cluster environment based on a parallel computing framework, a calculation module that executes a command is determined in advance using a calculation module list to determine whether the calculation module holds a buffer object used by the command. As a result, the buffer objects of other calculation modules are copied to the corresponding calculation module, thereby ensuring efficient memory management.

Meanwhile, the embodiments of the present invention can be embodied as computer readable codes on a computer readable recording medium. A computer-readable recording medium includes all kinds of recording apparatuses in which data that can be read by a computer system is stored.

Examples of the computer-readable recording medium include a ROM, a RAM, a CD-ROM, a magnetic tape, a floppy disk, an optical data storage device and the like, and also a carrier wave (for example, transmission via the Internet) . In addition, the computer-readable recording medium may be distributed over network-connected computer systems so that computer readable codes can be stored and executed in a distributed manner. In addition, functional programs, codes, and code segments for implementing the present invention can be easily deduced by programmers skilled in the art to which the present invention belongs.

Further, the embodiments described above are intended to illustrate the present invention, and the scope of the present invention is not limited to the specific embodiments.

220: host node
221: host thread
222: command scheduler
240: compute node
241: calculation module
242: command handler
244: module thread
260: network
280: memory management unit

Claims (8)

A host thread that executes a host program for the parallel computing framework and inserts commands related to the kernel program for the parallel computing framework into a command queue;
The command scheduler selects a command to be executed by scheduling the command queue, generates a request message requesting execution of the selected command, and transmits the generated request message via a network to a calculation node including a calculation module to execute the command. ; And
It is determined whether the first calculation module which is the calculation module to execute the selected command has a buffer object used by the selected command, and according to the determination result, the second calculation module which is a calculation module different from the first calculation module. A memory manager to copy a buffer object to the first calculation module; Host node of the parallel computing framework-based cluster system comprising a.
The apparatus of claim 1, wherein the memory management unit
A host node of the parallel computing framework-based cluster system, which generates and manages a calculation module list composed of calculation modules holding the latest data on the buffer objects for each buffer object.
The memory manager of claim 2, wherein the memory manager
And determining whether the first calculation module has the buffer object by using the calculation module list.
The memory manager of claim 3, wherein the memory manager
If the first calculation module is not in the calculation module list,
And copy the buffer object of the second computation module to the first computation module and add the first computation module to the computation module list.
The memory manager of claim 2, wherein the memory manager
When the buffer object is modified, the host of the parallel computing framework-based cluster system, after clearing the calculation module list of the buffer object, adds the calculation module that modified the buffer object to the calculation module list. Node.
The method of claim 1, wherein the command scheduler is
If the first calculation module does not hold the buffer object,
And host the request message after the buffer object of the second computing module is copied to the first computing module.
The method of claim 1, wherein the buffer object
Host node of the parallel computing framework based cluster system, characterized in that it is defined as an abstracted memory area used by the application for the parallel computing framework.
Executing a host program for the parallel computing framework and inserting a command associated with a kernel program for the parallel computing framework into a command queue;
Scheduling the command queue to select a command to be executed;
It is determined whether a first calculation module which is a calculation module to execute the selected command has a buffer object used by the selected command, and according to the determination result, a buffer of a second calculation module which is a calculation module different from the first calculation module. Copying an object to the first calculation module; And
Generating a request message requesting execution of the selected command and transmitting the generated request message to a calculation node including the first calculation module through a network; Memory management method of a parallel computing framework based cluster system comprising a.

Images