KR102123117B1 - Apparatus and method for predicting performance and power in multi-node system - Google Patents

Abstract

The present invention relates to an apparatus for predicting performance and power in a multi-node system and a method thereof. According to an embodiment, the apparatus for predicting performance and power comprises: an interface unit which receives parameters for a multi-node application, which includes a message passing interface (MPI) command, and parameters of hardware based on a multi-node, which performs the application; and a performance and power amount predictor which applies the received parameters to at least one of a performance prediction model and a power prediction model to predict a result value of at least one of performance information and a power amount according to the application performance in the multi-node.

Description

Apparatus and method for predicting performance and power in a multi-node system{APPARATUS AND METHOD FOR PREDICTING PERFORMANCE AND POWER IN MULTI-NODE SYSTEM}

The present invention relates to an apparatus and a method for predicting performance and power in a multi-node system, and more particularly, to a technical idea of predicting performance and power through data movement analysis in a multi-node system.

Parallel computing refers to a method of dividing a large operation into smaller operations and computing them in parallel at the same time.

In addition, in parallel computing, a set of computers connected to several computers and acting as one system is called a computer cluster. Clusters (multi-nodes) generally provide better performance and reliability than single computers (nodes), and are much more cost-effective than single computers that offer similar performance and reliability.

1 is a diagram for explaining an operation structure of a parallel program.

Referring to FIG. 1, reference numeral 100 denotes a result of comparing a serial processing method for computing data at a single node and a parallel processing method for computing data at multiple nodes over time.

According to reference numeral 100, the parallel processing method using multiple nodes can distribute and process a plurality of data P1 to P4 at the same time in multiple nodes (Node 1 to Node 4), so that the serial processing method using a single node Data operations can be performed at a faster rate.

2 is a diagram for explaining a network topology of multiple nodes.

2, reference numeral 200 denotes an example of a network topology type that physically connects elements (nodes) of a computer network.

According to reference numeral 200, the network topology is a ring type topology in which a plurality of nodes are connected in the form shown in (a) of reference numeral 200, and a mesh type connected in the form shown in (b). It can be divided into a topology, a star type topology connected in the form shown in (c), and a full connected topology connected in the form shown in (d).

In addition, the network topology includes a line type topology in which a plurality of nodes are connected in the form shown in (e) of reference numeral 200, a tree type topology connected in the form shown in (f), and ( It can also be divided into a bus type (Bus type) topology connected in the form shown in g).

Specifically, the network topology has a difference in a processing method for moving data according to a form, and also shows a difference in performance and power.

The existing technology for predicting performance and power for a node is difficult to apply to a currently widely used multi-node system for predicting performance and power for only a single node.

In addition, the existing technology has a problem in that accuracy and reliability of performance information and power output to predict performance and power are deteriorated without considering the shape of the topology in a multi-node system.

Korean Patent Publication No. 10-2010-7003117 "Proactive power management in parallel computers"

The present invention is to provide a performance and power prediction apparatus and method for easily predicting performance and power consumption for an application requiring operation of multiple nodes.

In addition, the present invention is to provide a performance and power prediction apparatus and method for predicting different performance and power change according to the type of topology when operating the same application.

In addition, the present invention intends to provide a performance and power predicting apparatus and a method for improving the accuracy and reliability of a predicted value for performance and power consumption by providing a predicted value of performance and power consumption in consideration of the topology type.

The apparatus for predicting performance and power according to an embodiment includes an interface unit for receiving an application for a multi-node including a message passing interface (MPI) command, and parameters of hardware based on a multi-node performing the application. And applying a parameter to at least one of the performance prediction model and the power prediction model to include a performance and power amount prediction unit for predicting a result value of at least one of performance information and power amount according to the application execution in the multiple nodes. .

According to one side, the parameter may include at least one of the number of nodes constituting the multi-node, network topology information of the multi-node, hardware information of each node constituting the multi-node, and idle power consumption information. have.

According to one side, at least one of the performance prediction model and the power prediction model receives the number of nodes constituting the multi-node and the network topology information of the multi-node as input, and performs performance information according to the application execution and It may be a model for predicting a result value of at least one of the amount of power.

According to one side, the performance and power prediction apparatus extracts and splits the MPI instruction included in the application, and predicts performance information according to the application performance based on the divided MPI instruction, and the divided A prediction model generator for generating the power prediction model predicting the amount of power according to the application execution based on the MPI instruction may be further included.

According to one side, the prediction model generating unit includes an MPI instruction included in the application as a serial instruction set for computing data at each node constituting the multi-node and a parallel instruction set for computing data at the multi-node. Can be divided.

According to one side, the prediction model generator derives the total execution time by summing up the sum of time spent in data operation for each node constituting the multi-node and the sum of time spent in data movement between the nodes. The performance prediction model can be generated.

According to one side, the time spent in calculating data for each node may be derived through a value reflecting a ratio between the amount of calculated data for each node and the data processing speed for each node.

According to one side, the time spent for data movement between each node may be derived through a value reflecting routing complexity in a ratio between the total amount of computed data of the multiple nodes and a predetermined bandwidth.

According to one side, the prediction model generator derives the total amount of power consumption through summation of power consumption according to data calculation for each node constituting the multi-node, and sum of power consumption according to data movement between the nodes. The power prediction model can be generated.

According to one side, the power consumption according to the data operation for each node is a value that reflects the total execution time in the multi-node to a previously measured static power consumption value, and the number of operations for each classification of the divided MPI instruction is reflected. It can be derived through the summation of the sum of dynamic power consumption for each star.

According to one side, the power consumption according to data movement between each node is a sum of values reflecting the unit power consumption value in a cable provided between each node constituting the multiple nodes and the time spent in data movement between the nodes. Can be derived through

A method of predicting performance and power according to an embodiment includes receiving an application parameter for a multi-node including a message passing interface (MPI) command from an interface unit and parameters of hardware based on a multi-node performing the application, and Predicting a result value of at least one of performance information and power according to the application performance in the multi-node by applying the received parameter from the performance and power prediction unit to any one of a performance prediction model and a power prediction model It may include.

According to one side, a performance and power prediction method is a prediction model generating unit extracting and dividing an MPI instruction included in the application from the prediction model generator, and predicting performance information according to the application execution based on the divided MPI instruction. And, based on the divided MPI command may further include generating the power prediction model for predicting the amount of power according to the application execution.

According to one embodiment, it is possible to easily predict performance and power consumption for an application requiring operation of multiple nodes.

According to an embodiment, when operating the same application, different performances and power changes can be predicted according to the shape of the topology.

According to an embodiment, by providing a predicted value of performance and power consumption in consideration of a topology form, accuracy and reliability of a predicted value for performance and power consumption can be improved.

1 is a diagram for explaining an operation structure of a parallel program.
2 is a diagram for explaining a network topology of multiple nodes.
3 is a diagram for describing a performance and power prediction apparatus according to an embodiment.
4 is a view for explaining a method for predicting performance and power according to an embodiment.

Specific structural or functional descriptions of the embodiments according to the concept of the present invention disclosed in this specification are exemplified only for the purpose of illustrating the embodiments according to the concept of the present invention, and the embodiments according to the concept of the present invention These can be implemented in various forms and are not limited to the embodiments described herein.

Embodiments according to the concept of the present invention can be applied to various changes and have various forms, so the embodiments will be illustrated in the drawings and described in detail herein. However, this is not intended to limit the embodiments according to the concept of the present invention to specific disclosure forms, and includes modifications, equivalents, or substitutes included in the spirit and scope of the present invention.

Terms such as first or second may be used to describe various components, but the components should not be limited by the terms. The above terms are only for the purpose of distinguishing one component from other components, for example, without departing from the scope of rights according to the concept of the present invention, the first component may be referred to as the second component, Similarly, the second component may also be referred to as the first component.

When an element is said to be "connected" to or "connected" to another component, it is understood that other components may be directly connected to or connected to the other component, but may exist in the middle. It should be. On the other hand, when a component is said to be "directly connected" or "directly connected" to another component, it should be understood that no other component exists in the middle. Expressions describing the relationship between the components, for example, "between" and "immediately between" or "directly adjacent to" should be interpreted similarly.

The terminology used herein is only used to describe specific embodiments and is not intended to limit the present invention. Singular expressions include plural expressions unless the context clearly indicates otherwise. In this specification, the terms "include" or "have" are intended to designate the presence of a feature, number, step, action, component, part, or combination thereof as described, one or more other features or numbers, It should be understood that the existence or addition possibilities of steps, actions, components, parts or combinations thereof are not excluded in advance.

Unless defined otherwise, all terms used herein, including technical or scientific terms, have the same meaning as commonly understood by a person skilled in the art to which the present invention pertains. Terms such as those defined in a commonly used dictionary should be interpreted as having meanings consistent with meanings in the context of related technologies, and should not be interpreted as ideal or excessively formal meanings unless explicitly defined herein. Does not.

Hereinafter, embodiments will be described in detail with reference to the accompanying drawings. However, the scope of the patent application is not limited or limited by these embodiments. The same reference numerals in each drawing denote the same members.

3 is a diagram for describing a performance and power prediction apparatus according to an embodiment.

Referring to FIG. 3, the apparatus for predicting performance and power according to an embodiment 300 can easily predict performance and power consumption for an application requiring operation of multiple nodes.

In addition, the performance and power prediction apparatus 300 may predict different performance and power changes according to the topology type when operating the same application.

In addition, the performance and power predicting apparatus 300 may improve the accuracy and reliability of the predicted values for performance and power consumption by providing the predicted values for performance and power consumption in consideration of the topology type.

To this end, the performance and power prediction apparatus 300 may include an interface unit 310, a prediction model generator 320, and a performance and power amount prediction unit 330.

Specifically, the interface unit 310 according to an embodiment may receive parameters for a multi-node application and a multi-node-based hardware for performing an application for an application including an MPI (Message Passing Interface) command.

Here, MPI means a standard that describes the exchange of information in distributed and parallel processing.

More specifically, MPI refers to a programming model that implements communication between processes only by sending (sending) and receiving (receiving) messages in a multi-node system composed of processes having their own memory locally.

The parameters may be hardware parameters of a multi-node system aimed at predicting performance and power consumption.

In other words, the performance and power prediction apparatus 300 according to an embodiment generates a prediction model through a multi-node application that includes an MPI instruction through multi-node A, and multi-nodes aiming at prediction of performance and power consumption. By applying the hardware parameters for B as the input of the generated prediction model, it is possible to predict the performance and power consumption according to application use in the multi-node B.

According to one side, the parameter may include at least one of the number of nodes constituting the multiple nodes, the network topology information of the multiple nodes, the hardware information of each node constituting the multiple nodes, and the idle power consumption information.

In addition, the parameters may further include data efficiency information according to the network topology, but are not limited to the above-described examples, and various parameters that may be derived from a multi-node system may be included.

For example, the network topology information of multiple nodes includes information on the network topology type, and the topology types are ring type, mesh type, star type, and full connected type. ), a line type (Line type), a tree type (Tree type) and may include at least one of a bus type (Bus type), but the topology type is not limited to the above-described examples, and may include various connection types. .

Next, the performance and power amount predicting unit 330 according to an embodiment applies at least one of the performance prediction model and the power prediction model to the received parameters to perform at least one of performance information and power amount according to application execution in multiple nodes. One result can be predicted.

According to one side, at least one of the performance prediction model and the power prediction model receives at least one of the number of nodes constituting the multi-nodes and the network topology information of the multi-nodes as inputs, and the performance information and the amount of power according to application execution. It may be a model for predicting the result.

In other words, the performance prediction model and the power prediction model according to an embodiment include information on the number of nodes constituting the multi-node B, which is a parameter related to the multi-node B, and the network topology of the multi-node B. It may be a model that receives information as input and predicts performance and power consumption according to application use in multiple Node Bs.

More specifically, the performance prediction model and the power prediction model may be a model that infers a data movement amount and a calculation amount for each node according to the number of nodes, and predicts total execution time and power consumption in multiple nodes according to network topology information.

In other words, the performance prediction model and the power prediction model can predict the execution time and power consumption of instructions according to the network topology by analyzing the data movement amount when a specific application is executed in multiple nodes.

Meanwhile, the prediction model generator 320 extracts and splits the MPI instructions included in the application, and predicts performance information according to application performance based on the divided MPI instructions, and based on the divided MPI instructions. A power prediction model for predicting the amount of power according to application execution may be generated.

More specifically, the prediction model generator 320 may extract an MPI instruction while an application operation is performed, and the MPI instruction may include information regarding which processor (node) to which processor (node) is requesting data. Size information of requested data is specified.

In addition, MPI commands can be requested for one-to-many, many-to-one communication as well as real-time extraction when the application is running. When the MPI command is called when the application is executed, the MPI commands are parsed to extract the data movement amount. Can be.

According to one side, the prediction model generation unit 320 includes MPI instructions included in an application as a serial instruction set for computing data at each node constituting multiple nodes and a parallel instruction set for computing data at multiple nodes. Can be divided.

More specifically, the prediction model generator 320 may check the order of the performed instructions and analyze the operation in hardware. At this time, the prediction model generator 320 may classify instruction sections processed in multiple nodes based on MPI instructions called between instructions.

Through this, the prediction model generator 320 can divide the instruction sequence into a serial processing section in a single node and a parallel processing section in multiple nodes. As described above, the MPI command includes a starting node, a target node, Information about the amount of data is included, so you can grasp the amount of data transmitted.

That is, the prediction model generator 320 may grasp the type of operation using the divided instruction set, and may estimate the amount of operation and the amount of data for each node using the type of operation.

According to one side, the prediction model generator 320 may divide the MPI instruction into a serial instruction set and a parallel instruction set, as well as a data movement instruction set related to data movement between nodes, but is not limited to the above-described example. And it can be divided into a variety of criteria required for power consumption prediction.

According to one side, the prediction model generator 320 is the sum of the time (T _{multi -node} ) spent in data calculation for each node constituting multiple nodes, and the time spent in data movement between each node (T _{data - trasnsfer).} ) Can generate a performance prediction model that derives the _total execution time (T _total ) through the summation of ).

In other words, the performance prediction model according to an embodiment may be expressed by Equation 1 below.

[Equation 1]

According to one side, in Equation 1, the time (T _{data -trasnsfer} ) spent on data calculation for each node is the following Equation 2, which derives the largest value among the execution time values (T _{single -node} ) of a single node. Can be derived through

[Equation 2]

According to one side, the time (T _{multi -node} ) spent on data calculation for each node reflects the ratio between the amount of computation data (N _single-data ) for each node and the data processing speed (F _{single -node} ) for each node. It can be derived from a value.

In other words, in Equation 2, the operation time value (T _single-node ) of a single node for deriving the time (T _{multi -node} ) spent on data calculation for each node may be derived through Equation 3 below. have.

[Equation 3]

For example, the data processing speed (F _{single -node} ) for each node may be determined through at least one information of hardware performance, memory structure, and operating frequency of each node, which may be utilized as described in hardware performance of each node or It can be obtained through measurement of target hardware. Further, the hardware of each node may be a processor corresponding to each node.

On the other hand, the number of processors that can be utilized may vary depending on the number of nodes, and a difference may occur in parallelism, which may also affect the performance of parallel computation in multiple nodes.

The number of nodes and the parallelism that can be utilized to the maximum have a proportional relationship, and as the parallelism increases, the throughput per node may decrease.

According to one side, the total amount of operation data (N _data ) of multiple nodes may vary depending on application characteristics, and may be derived through code analysis of MPI instructions.

In addition, the total operation data amount N _data of multiple nodes may be composed of several parallel operations (N _{op_i} , where i is a natural number of 1 or more), and each individual operation may be performed by being divided into multiple nodes.

That is, the amount of operation data (N _single-data ) for each node may be derived through Equation 4 below, which divides several parallel operations (N _{op _i} ) by the number (n) of nodes on which multiple operations are performed. .

[Equation 4]

According to one side, the time spent in data movement between each node (T _{data - trasnsfer} ) reflects the routing complexity (C _R ) in the ratio between the total amount of computed data (N _data ) and the preset bandwidth (B) of multiple nodes. It can be derived from a value.

In other words, the time (T _{data - trasnsfer} ) spent on data movement between each node may be derived through Equation 5 below.

[Equation 5]

Here, square brackets mean a Gaussian symbol, and N _basic indicates a unit size when data is moved.

More specifically, in Equation 5, when the unit size (N _basic ) is 64 bytes, data is moved in units of 64 bytes in the same way when sending 10 bytes or sending 40 bytes, so a Gaussian value is added to match the unit size (N _basic ). Can be.

Meanwhile, in Equation 5, the routing complexity (C _R ) may be determined according to a routing scheme.

For example, the routing scheme can be determined based on the multi-node network topology type.

In addition, the routing complexity (C _R ) can be derived through Equation 6 below, which calculates the minimum hop count according to the routing scheme.

[Equation 6]

According to one side, the prediction model generating unit 320 is the sum of the power consumption (W _{single -node} ) according to data operation for each node constituting multiple nodes, and the power consumption (W _{data -transfer)} according to data movement between each node. ) Can generate a power prediction model that derives the _total power consumption (W _total ) through the summation of ).

In other words, the power prediction model according to an embodiment may be expressed by Equation 7 below.

[Equation 7]

According to one side, the power consumption (W _{single -node} ) according to the data calculation for each node is a value that reflects the _total execution time (T _total ) at multiple nodes to the previously measured static power consumption value (P _static ), and is divided. It can be derived through summation of the sum of dynamic power consumption (W _compute ) for each node in which the number of operations (N _inst ) for each classification of the MPI instruction is reflected.

In other words, power consumption (W _{single -node} ) according to data calculation for each node may be derived through Equation 8 below.

[Equation 8]

Here, the static power consumption (P _static ) may be derived from a value described through each node or a value measured through experiments.

Meanwhile, in Equation 8, the dynamic power consumption (W _compute ) for each node is a value reflecting the difference in power consumed when calculating data at each node, and the _static power consumption of the target node (W _static ) to measure power It can be derived through Equation 9 below, which reflects the dynamic power consumption of the reference node (W _stdcompute ) in the ratio between the static power consumption of the reference node (W _stdstatic ).

[Equation 9]

According to one side, in Equation 9, the reference node may mean a node (a single computer) capable of measuring power. In addition, a ratio between the _static power consumption (W _static ) and the static power consumption (W _stdstatic ) of the reference node may be used to obtain a power value of a node that cannot measure power.

For example, in the case of the CPU, the static power consumption (W _static ) and the static power consumption (W _stdstatic ) of the reference node may be determined through thermal design power (TDP) provided by the manufacturer.

Meanwhile, in Equation 9, the dynamic power consumption of the reference node (W _stdcompute ) reflects the following Equation 10 reflecting the reference power consumption (P _stdInst ) according to the classification of the instruction in the number of operations (N _inst ) by classification of the divided MPI instruction. Can be derived through

[Equation 10]

Here, i may be a natural number of 1 or more.

For example, the number of operations per classification (N _inst ) may be the number of operations per category of instructions performed in each node, and the instructions may be MPI instructions divided through the prediction model generator 320.

Further, the instructions may be an instruction classified as at least one of an integer data operation instruction, a floating-point data operation instruction, and a branch control instruction.

Meanwhile, the reference power consumption P _stdInst is power consumption according to each instruction classification, and may be derived through experimentation.

According to one side, the power consumption (W _{data -transfer} ) due to data movement between each node is consumed for data movement between each node in the unit power consumption (P _transfer ) value in a cable provided between each node constituting multiple nodes. It can be derived by summing the values reflecting the time (T _trasnsfer ).

In other words, power consumption (W _{data -transfer} ) according to data movement between nodes may be derived through Equation 11 below.

[Equation 11]

Here, i may be a natural number of 1 or more.

Specifically, according to Equation 11, the power consumption (W _{data-transfer} ) of data movement between each node is the time (T _transfer ) of data movement between each node and the unit power consumption (P _transfer ) of the cable. It can be expressed as the sum of products. In addition, the unit power consumption (P _transfer ) in the cable depends on the cable connected between the nodes, and may be determined by a value provided by the manufacturer.

According to one side, the time (T _transfer ) spent on data movement between each node may be derived through Equation 5 above. That is, a value of power consumption (W _{data -transfer} ) according to data movement between each node derived through Equation 11 may be derived according to a network topology type of multiple nodes.

4 is a view for explaining a method for predicting performance and power according to an embodiment.

In other words, FIG. 4 relates to a method for predicting performance and power using a performance and power predicting apparatus according to an embodiment described with reference to FIG. 3, which is later described with reference to FIG. The description will be omitted.

Referring to FIG. 4, in step 410, the performance and power prediction method according to an embodiment may receive an application for a multi-node including a message passing interface (MPI) command from the interface unit.

According to one side, in step 420, the performance and power prediction method according to an embodiment extracts and splits the MPI instruction included in the application from the prediction model generator, and performs the performance according to the application execution based on the divided MPI instruction. The performance prediction model for predicting information and the power prediction model for predicting the amount of power according to the application execution may be generated based on the divided MPI instruction.

Next, in step 430, the performance and power prediction method according to an embodiment may receive parameters of hardware based on multi-nodes performing an application for multiple nodes in the interface unit.

Next, in step 440, the performance and power prediction method according to an embodiment applies the received parameters from the performance and power estimation unit to any one of the performance prediction model and the power prediction model to apply the application in the multi-node. At least one result value among performance information and power amount according to performance may be predicted.

As a result, using the present invention, it is possible to easily predict performance and power consumption for an application requiring operation of multiple nodes.

Also, when operating the same application, different performance and power changes can be predicted according to the topology connection type.

In addition, by providing the predicted values of performance and power consumption in consideration of the connection type of the topology, it is possible to improve the accuracy and reliability of the predicted values for performance and power consumption.

The device described above may be implemented with hardware components, software components, and/or combinations of hardware components and software components. For example, the devices and components described in the embodiments include, for example, processors, controllers, arithmetic logic units (ALUs), digital signal processors (micro signal processors), microcomputers, field programmable gate arrays (FPGAs). , A programmable logic unit (PLU), microprocessor, or any other device capable of executing and responding to instructions, may be implemented using one or more general purpose computers or special purpose computers. The processing device may run an operating system (OS) and one or more software applications running on the operating system. In addition, the processing device may access, store, manipulate, process, and generate data in response to the execution of the software. For convenience of understanding, a processing device may be described as one being used, but a person having ordinary skill in the art, the processing device may include a plurality of processing elements and/or a plurality of types of processing elements. It can be seen that may include. For example, the processing device may include a plurality of processors or a processor and a controller. In addition, other processing configurations, such as parallel processors, are possible.

The software may include a computer program, code, instruction, or a combination of one or more of these, and configure the processing device to operate as desired, or process independently or collectively You can command the device. Software and/or data may be interpreted by a processing device, or to provide instructions or data to a processing device, of any type of machine, component, physical device, virtual equipment, computer storage medium or device. , Or may be permanently or temporarily embodied in the transmitted signal wave. The software may be distributed over networked computer systems, and stored or executed in a distributed manner. Software and data may be stored in one or more computer-readable recording media.

The method according to the embodiment may be implemented in the form of program instructions that can be executed through various computer means and recorded on a computer-readable medium. The computer-readable medium may include program instructions, data files, data structures, or the like alone or in combination. The program instructions recorded in the medium may be specially designed and configured for the embodiments or may be known and usable by those skilled in computer software. Examples of computer-readable recording media include magnetic media such as hard disks, floppy disks, and magnetic tapes, optical media such as CD-ROMs, DVDs, and magnetic media such as floptical disks. Includes hardware devices specifically configured to store and execute program instructions such as magneto-optical media, and ROM, RAM, flash memory, and the like. Examples of program instructions include high-level language code that can be executed by a computer using an interpreter, etc., as well as machine language codes produced by a compiler. The hardware device described above may be configured to operate as one or more software modules to perform the operations of the embodiments, and vice versa.

Although the embodiments have been described by the limited drawings as described above, a person skilled in the art can make various modifications and variations from the above description. For example, the described techniques are performed in a different order than the described method, and/or the components of the described system, structure, device, circuit, etc. are combined or combined in a different form from the described method, or other components Alternatively, even if replaced or substituted by equivalents, appropriate results can be achieved.

Therefore, other implementations, other embodiments, and equivalents to the claims are also within the scope of the following claims.

300: performance and power prediction device 310: interface
320: prediction model generation unit 330: performance and power amount prediction unit

Claims (13)

An interface unit for receiving a parameter for a multi-node application including MPI (Message Passing Interface) commands and multi-node-based hardware for performing the application;
A performance and power amount prediction unit for predicting at least one result of performance information and power amount according to the application performance in the multi-node by applying the received parameter to at least one of a performance prediction model and a power prediction model;
The performance prediction model for extracting and partitioning MPI instructions included in the application, and predicting performance information according to the application performance based on the divided MPI instruction, and the application performance based on the divided MPI instruction. Prediction model generation unit for generating the power prediction model for predicting the amount of power
Including,
The prediction model generator
The MPI instruction included in the application is divided into a serial instruction set for computing data at each node constituting the multi-node and a parallel instruction set for computing data at the multi-node.
Multi-node performance and power prediction device. According to claim 1,
The above parameters
Including at least one of the number of nodes constituting the multi-node, network topology information of the multi-node, hardware information of each node constituting the multi-node, and idle power consumption information.
Multi-node performance and power prediction device. According to claim 1,
At least one of the performance prediction model and the power prediction model is
A model that receives the number of nodes constituting the multi-node and the network topology information of the multi-node as an input and predicts a result value of at least one of performance information and power according to the application execution
Multi-node performance and power prediction device. delete delete An interface unit for receiving a parameter for a multi-node application including MPI (Message Passing Interface) commands and multi-node-based hardware for performing the application;
A performance and power amount prediction unit for predicting at least one result value of performance information and power amount according to the application execution in the multi-node by applying the received parameter to at least one of a performance prediction model and a power prediction model;
The performance prediction model for extracting and dividing the MPI instruction included in the application, and predicting performance information according to the application execution based on the divided MPI instruction, and the application execution based on the divided MPI instruction. Prediction model generation unit for generating the power prediction model for predicting the amount of power
Including,
The prediction model generator
Generating the performance prediction model that derives the total execution time through the sum of the time spent in data operation for each node constituting the multi-node and the sum of time spent in data movement between each node
Multi-node performance and power prediction device. The method of claim 6,
The time spent in data calculation for each node is
Derived through a value reflecting the ratio between the calculation data amount for each node and the data processing speed for each node
Multi-node performance and power prediction device. The method of claim 6,
The time spent for data movement between each node is
Derived through a value reflecting the routing complexity in the ratio between the total amount of computed data of the multi-nodes and a predetermined bandwidth
Multi-node performance and power prediction device. An interface unit for receiving a parameter for a multi-node application including MPI (Message Passing Interface) commands and multi-node-based hardware for performing the application;
A performance and power amount prediction unit for predicting at least one result value of performance information and power amount according to the application execution in the multi-node by applying the received parameter to at least one of a performance prediction model and a power prediction model;
The performance prediction model for extracting and dividing the MPI instruction included in the application, and predicting performance information according to the application execution based on the divided MPI instruction, and the application execution based on the divided MPI instruction. Prediction model generation unit for generating the power prediction model for predicting the amount of power
Including,
The prediction model generator
Generating the power prediction model to derive the total amount of power consumption by summing the sum of power consumption according to data operation for each node constituting the multi-node and the sum of power consumption according to data movement between the nodes
Multi-node performance and power prediction device. The method of claim 9,
The power consumption according to the data calculation for each node is
Derived from the sum of the sum of the dynamic power consumption for each node, which reflects the total execution time in the multi-node and the calculated number of operations of each segmented MPI instruction in the previously measured static power consumption value.
Multi-node performance and power prediction device. The method of claim 9,
The power consumption according to data movement between the nodes is
It is derived through the sum of values that reflect the time spent for data movement between each node and the unit power consumption value in the cable provided between each node constituting the multi-node.
Multi-node performance and power prediction device. Receiving a parameter for a multi-node application including MPI (Message Passing Interface) commands from the interface unit and multi-node-based hardware for performing the application;
Predicting a result value of at least one of performance information and power according to the application performance in the multi-node by applying the received parameter from the performance and power prediction unit to any one of a performance prediction model and a power prediction model And
The prediction model generation unit extracts and splits the MPI instruction included in the application, and predicts performance information according to the application performance based on the divided MPI instruction, and based on the divided MPI instruction. Generating the power prediction model to predict the amount of power according to the application execution
Including,
The prediction model generator
The MPI instruction included in the application is divided into a serial instruction set for computing data at each node constituting the multi-node and a parallel instruction set for computing data at the multi-node.
A method for predicting performance and power of multiple nodes. delete

Images