KR20250021054A - Data processing system including remote processing device and operation method thereof - Google Patents

Abstract

A data processing system according to the present invention comprises a host executing a program for processing a data set; and a remote processing device connected to the host via an interface, wherein the host comprises a profile control circuit for generating a profile corresponding to a function called when the program is executed; a profile database for storing execution locations of functions corresponding to the profiles; and a policy execution circuit for transmitting a function called when the program is executed to the host or the remote processing device by referring to the profile database.

Description

DATA PROCESSING SYSTEM INCLUDING REMOTE PROCESSING DEVICE AND OPERATION METHOD THEREOF

The present technology relates to a data processing system including a remote processing device and an operating method thereof, and relates to a data processing system and an operating method thereof in which a program for processing a large amount of data is distributed and processed between a host and a remote processing device.

In order for a data processing system to execute programs that process large amounts of data, such as clustering programs, the main memory capacity must be very large, and the communication cost between the CPU and memory also increases significantly.

Accordingly, a distributed processing technology is being proposed by combining an NDP (Near Data Processing) device that includes a large-capacity remote memory connected to a host.

However, compared to accessing main memory, accessing remote memory has relatively higher latency and bandwidth is limited due to interface limitations, which limits the efficiency of distributed processing.

US 2006/0047794 A1 KR 10-2016-0081815 A

This technology provides a technology for profiling applications that process large amounts of data and efficiently distributing and processing the applications on hosts and remote processing devices by reflecting the profiling results.

A data processing system according to one embodiment of the present invention comprises: a host executing a program for processing a data set; and a remote processing device connected to the host via an interface, wherein the host comprises: a profile control circuit for generating a profile corresponding to a function called when the program is executed; a profile database for storing execution locations of functions corresponding to the profiles; and a policy execution circuit for transferring a function called when the program is executed to the host or the remote processing device by referring to the profile database.

A method of operating a data processing system according to one embodiment of the present invention comprises: a host executing a program for processing a data set; and a remote processing device connected to the host via an interface, the method comprising: a profile generation step of storing a plurality of profiles corresponding to a plurality of functions included in a program while executing the program using a synthetic data set having a smaller size than the data set; and an optimization step of determining an execution location of a function corresponding to a plurality of profiles as either the host or the remote processing device while executing the program using the synthetic data set.

The data processing system by this technology performs profiling on functions included in a program using the program context, applies an optimization algorithm to the profiling results, and automatically distributes and processes the program between the host and a remote processing device, thereby enabling efficient system utilization.

Figure 1 is a block diagram showing a data processing system according to one embodiment of the present invention.
FIG. 2 is an explanatory diagram showing the profile operation of a data processing system according to one embodiment of the present invention.
Figure 3 is a table showing the structure of a profile database according to one embodiment of the present invention.
Figure 4 is a flowchart showing the optimization operation of a data processing system according to one embodiment of the present invention.
FIG. 5 is an explanatory diagram showing the optimization operation of a data processing system according to one embodiment of the present invention.

Hereinafter, embodiments of the present invention will be described with reference to the attached drawings.

Figure 1 is a block diagram showing a data processing system (1) according to one embodiment of the present invention.

The data processing system (1) includes a host (10) and a remote processing device (20).

The host (10) includes one or more processors (11) and local memory (12). The processor (11) includes a conventional CPU and the local memory (12) may be referred to as main memory.

The host (10) further includes a profile control circuit (110), a profile database (120), a policy execution circuit (130), and a policy decision circuit (140).

The operations of the profile control circuit (110), the profile database (120), the policy execution circuit (130), and the policy decision circuit (140) are specifically described below.

The remote processing device (20) supports near data processing (NDP) and may be referred to as an NDP device.

The remote processing device (20) includes one or more computational circuits (21) and a remote memory (22), and the computational circuit (21) performs an NDP function using data stored in the remote memory (22).

The technology itself regarding a remote processing device (20) that performs an NDP function by having an operation circuit (21) and a remote memory (22) is a well-known technology.

The host (10) and the remote processing device (20) can be connected through various interface technologies.

In this embodiment, it is assumed that the host (10) and the remote processing device (20) are connected through an interface that supports CXL (Compute Express Link) technology according to the Type-2 standard.

Since the CXL technology itself by the Type-2 standard is well known, a detailed description of the interface technology is omitted.

When supporting CXL technology, the processor (11) of the host (10) can access remote memory (22) in the same manner as reading/writing local memory (12).

Through this interface technology, the host (10) can offload functions and data to remote memory and receive the processing results of the functions.

The data processing system (1) according to the present technology can process some functions of a program that processes a large amount of data in the host (10) and some functions in the remote processing device (20).

Below, the present invention is disclosed using a clustering program as an example.

Clustering programs are a technique for labeling data stored in a data set in an unsupervised manner.

Since there are various clustering programs depending on the clustering algorithm used, it is difficult to apply the distributed processing method optimized for a specific program to other programs.

In this embodiment, the host (10) sequentially executes a profile operation for various functions included in the program and an optimization operation for allocating the functions to targets to realize automated allocation for functions called when the program is executed.

At this time, the target corresponds to either the host (10) or the remote processing device (20).

In this embodiment, the program context associated with the call path of the function is used to generate a profile corresponding to the function, and a genetic algorithm is applied for optimization, but the scope of the present invention is not limited thereto.

In the profile and optimization operations, the clustering program is run using a small synthetic data set rather than the actual data set. The synthetic data set can be generated randomly or by sampling a portion of the actual data set.

First, initiate the profile operation.

During the program execution process, various functions are called.

In this embodiment, profiling operations are performed using the program context corresponding to the function, rather than the individual function itself.

This creates a profile that reflects various contextual information in which the function is called.

In this embodiment, the program context reflects a series of function calls stacked in stack memory.

More specifically, in this embodiment, a hashed value of an array of return addresses of a series of functions stacked in the stack memory is used as a program context corresponding to the currently called function, and the hashed value has, for example, a 32-bit value.

Accordingly, even if it is the same function, it can have different program contexts depending on the context in which it is called.

Accordingly, profiles and program contexts can be used interchangeably below.

The profile control circuit (110) determines a profile corresponding to a function. In this embodiment, as described above, a profile corresponding to a function is generated by hashing an array of return addresses of functions stacked in the stack memory at the time the function is called.

The profile database (120) is a database that stores profiles created in response to called functions.

In this embodiment, the profile database (120) stores a profile corresponding to a called function, but stores it in association with the type of the called function.

In this embodiment, the types of functions are classified as operation functions or memory functions. For example, a memory function is a function that controls memory, such as an operation to allocate memory, and an operation function is a function other than this.

The operation functions may include thread allocation functions and other general functions that perform general operations.

When saving a new profile, you can capture the program context and save the relationships between profiles.

For example, if a memory function is called during the process of performing an operation function, the operation function and the memory function can be viewed as being related.

Accordingly, after a profile corresponding to each function is generated, the relationship between the two profiles can be stored together, and by querying either profile, the other related profile can be known.

Figure 2 is an explanatory diagram showing the profile operation of a data processing system (1) according to one embodiment of the present invention.

Figure 2 shows that functions from f0 to f6 are called sequentially, and the scope of each function is indicated by an arrow. For a function that does not have a corresponding arrow, it shows that the next function is called after the operation is completed.

The dotted rectangle indicates the return address corresponding to the function, but the return address may be different depending on the time of calling for the same function. For example, f ₅ and f ₅ ', which are indicated with dotted rectangles, indicate that they were called at different times and have different return addresses.

In Figure 2, f ₁ and f ₅ correspond to memory allocation functions, f ₄ corresponds to a thread creation function, and the rest correspond to general functions. That is, f ₄ corresponds to a memory function, and the rest correspond to operation functions.

Figure 2 shows the time at which functions are called and the scope of the life of the called function. For example, f ₀ is called and ends after the operation of f ₆ is finished. In the case of a function without an arrow, it indicates that it returns after the call and before the next function is called.

In this embodiment, the return address array represents an array containing the return address of the currently executing procedure or function stored in stack memory.

The return address array (A(f _i )) corresponding to the function (f _i ) called in Fig. 2 corresponds to the array of return addresses sequentially stored in the stack memory when the function (f _i ) is called.

First, when f ₀ is called, the return address of f ₀ is stored in the stack memory, and for subsequent function calls, the return addresses of the corresponding functions are sequentially stored in the stack memory.

For example, when f ₁ is called, the return address of f ₁ is stored after the return address of f ₀ , when f ₂ is called after f ₁ returns, the return address of f ₂ is stored after the return address of f ₀ , and when f ₃ is called, the return addresses of f ₂ and f ₃ are stored sequentially after the return address of f ₀ .

In this embodiment _, the profile corresponding to the function (f _i ) corresponds to the hash value of the return address array (A(f _i )) corresponding to the function (f i ).

At this time, the hash function takes an array of return addresses as input. For example, if the hash function is expressed as HASH(), the profile corresponding to f0 can be expressed as C(f ₀ ) = HASH(<addr _f0 >), and the profile corresponding to f1 can be expressed as C(f ₁ ) = HASH(<addr _f0 , addr _f1 >).

As the clustering program is executed, functions are called sequentially, and the profile control circuit (110) stores the profile generated by the above principle in the profile database (120).

FIG. 3 is a diagram showing the structure of a profile database (120) according to one embodiment of the present invention.

The profile database (120) includes a profile field, a profile type field, a related profile field, and an execution location field.

The profile field sequentially stores the profiles generated as described above.

The Profile Type field indicates the type of profile generated, which is classified as either Operational or Memory, depending on the type of function it corresponds to.

For example, the profile corresponding to the thread creation function of the aforementioned general function is displayed as operation, and the profile corresponding to the memory allocation function is displayed as memory.

The Related Profile field stores the profiles associated with the current profile.

For example, if the type of the current profile is memory, the profile corresponding to the operation function that called the corresponding memory allocation function is stored.

For example, in Figure 2, the memory allocation function f ₁ is called from the general function f ₀ .

When this is applied to Fig. 3, assuming that the profile of f ₀ is 3FE033B5 and the profile of f ₁ is 5964F1CB, 3FE033B5 can be stored as an associated profile of the profile of f _1. In addition, 5964F1CB can be additionally stored as an associated profile of the profile of f ₀ .

The execution location field stores the location where the function corresponding to the profile is to be executed. For example, if the function corresponding to the profile is to be executed on the host (10), 0 can be stored, and if the corresponding function is to be executed on a remote processing device (20), 1 can be stored.

The type of profile and the related profile can be used in the optimization operation, and the value of the execution location is ultimately determined through the optimization operation.

In this way, a profile database (120) is created using all functions called while executing the clustering program, and then an optimization operation is performed.

Below we describe the optimization operation.

The policy decision circuit (140) determines the execution location of each profile stored in the profile database (120) as the host (10) or remote processing device (20), and the policy execution circuit (130) instructs the host (10) or remote processing device (20) to execute a function according to the execution location of the determined profile.

The policy decision circuit (140) controls the optimization operation to determine the execution location corresponding to each profile according to the optimization algorithm.

Although various algorithms can be used as optimization algorithms, the following example discloses an embodiment using a genetic algorithm as an example.

FIG. 4 is a flowchart showing the optimization operation of a data processing system (1) according to one embodiment of the present invention, and FIG. 5 is a corresponding explanatory diagram.

First, a number of Nth generation chromosome vectors corresponding to the profiles included in the profile database (120) are generated (S10).

At this time, the variable N representing the generation is 0 or a natural number and is initialized to 0. Hereinafter, the 0th generation chromosome vector can be referred to as the initial chromosome vector.

In this embodiment, the number of N generation chromosome vectors is the same, and Fig. 5 exemplifies that four chromosome vectors are generated per generation.

Each element of the initial chromosome vector corresponds to one of the profiles included in the profile database (120), and its value indicates the execution location of the function corresponding to the profile.

For example, if the execution location is the host, the value is 0, and if the execution location is a remote processing device, the value is 1.

The initial chromosome vector can be set to a completely random value, but it may be impossible to obtain a converged chromosome vector as a result of the operation of the genetic algorithm, or it may take too long to converge.

In this embodiment, a bootstrapping technique is applied to prevent this phenomenon. To this end, in this embodiment, the value of the initial chromosome vector is initialized to a random value, and the value of the initial chromosome vector is changed by referring to the relevant profile.

For example, if the profile type is memory, change the values of the initial chromosome vector to have the same values as the related profile.

As illustrated in Fig. 3, since profile "5964F1CB" whose profile type is memory is related to profile "3FE033B5", the value corresponding to profile "5964F1CB" in the initial chromosome vector is changed to have the same value as the value corresponding to profile "3FE033B5".

Figure 5 illustrates one initial chromosome vector generated when it is assumed that the number of profiles stored in the profile database (120) is 8.

Next, the value of the execution location field of Fig. 3 is set using each Nth generation chromosome vector, and accordingly, a clustering program is executed using a synthetic data set, and then a score is derived by a pre-specified index (S20).

Predefined indicators can be selected in relation to the performance expected for the data processing system (1). For example, IPC (instruction per cycle), execution time, etc. can be used as predefined indicators.

Figure 5 shows four scores calculated corresponding to four Nth generation chromosome vectors X1, X2, X3, and X4.

Afterwards, it is determined whether N is greater than the maximum number of generations (S30). The maximum number of generations is determined in advance to be sufficiently large so that the chromosome vector converges. Since the convergence condition of the chromosome vector can be determined in various ways by applying a conventional genetic algorithm, a specific disclosure is omitted.

If N is not greater than the maximum number of generations, a pair of parent chromosome vectors for generating a plurality of next generation chromosome vectors is determined (S40).

A parent chromosome vector is stochastically selected from a number of N-th generation chromosome vectors, and the probability that any one N-th generation chromosome vector is selected is proportional to the score corresponding to any one N-th generation chromosome vector. At this time, each chromosome vector can be selected repeatedly.

Figure 5 shows four parent chromosome vectors P1, P2, P3, and P4 randomly selected from four Nth generation chromosome vectors.

Parent chromosome vectors sequentially generate parent chromosome vector pairs, and the next generation chromosome vector is generated from each parent chromosome vector pair.

A plurality of (N+1) generation chromosome vectors are generated from parent chromosome vector pairs, and in this process, crossover and mutation techniques using a genetic algorithm can be applied (S50). Since the crossover and mutation techniques themselves are conventional techniques, their description is omitted.

Figure 5 shows two (N+1)-th generation chromosome vectors Y1, Y2 generated from the first parent chromosome vector pair (P1, P2), and two (N+1)-th generation chromosome vectors Y3, Y4 generated from the second parent chromosome vector pair (P3, P4).

Afterwards, the variable (N) representing the generation is increased by 1 (S60), and the process returns to step (S20) and repeats the aforementioned operation.

If N is greater than the maximum number of generations, the chromosome vector corresponding to the highest score among multiple Nth generation chromosome vectors is selected as the optimal chromosome vector to finally determine the execution location (S70).

Figure 5 shows that the execution location corresponding to each profile is determined by the host (10) or remote processing device (20) according to the optimal chromosome vector.

When the aforementioned profile operation and optimization operation are completed, the original data set and clustering algorithm are applied to the data processing system (1) to perform the clustering operation.

The scope of the present invention is not limited to the above disclosure. The scope of the present invention should be interpreted based on the scope literally described in the claims and the equivalent scope thereof.

1: Data processing system
10: Host
11: Processor
12: Local memory, main memory
110: Profile control circuit
130: Policy execution circuit
140: Policy Decision Circuit
120: Profile Database
20: Remote Processing Unit
21: Operation circuit
22: Remote Memory

Claims (15)

a host that runs a program that processes the data set; and
A remote processing device connected to the host and interface above
Including, but not limited to,
The above host is
A profile control circuit that generates a profile corresponding to a function called when the above program is executed;
A profile database that stores the execution locations of functions corresponding to the profile; and
A policy execution circuit that refers to the above profile database and transmits a function called when the above program is executed to the host or the remote processing device.
A data processing system comprising: A data processing system according to claim 1, wherein the profile control circuit generates a profile corresponding to a function from an array of return addresses of the function stored in a stack memory when the function is called. In claim 1, a policy decision circuit is further included that controls an optimization operation for determining an execution location of the profile database while executing the program using a synthetic data set of a smaller size than the data set,
A data processing system in which the above optimization operation is executed after performing a profile operation of storing a profile corresponding to a function included in the program in the profile database while executing the program using the above synthetic data set. In claim 3, the policy decision circuit is a data processing system that performs the optimization operation using a genetic algorithm, and generates a plurality of initial chromosome vectors including a plurality of elements corresponding to a plurality of profiles included in the profile database during the optimization operation. A data processing system according to claim 4, wherein the policy decision circuit arbitrarily sets the value of each element of the initial chromosome vector to a value corresponding to either the host or the remote processing device. A data processing system according to claim 5, wherein the policy decision circuit adjusts a value of the initial chromosome vector so that, if any one element of the initial chromosome vector is a profile corresponding to a memory operation, the value has the same value as an element corresponding to a profile corresponding to an operation operation associated with the memory operation. In claim 4, the policy decision circuit determines a plurality of parent chromosome vector pairs according to a probabilistically selected order from a plurality of N-th generation chromosome vectors, and applies crossover and mutation techniques to each of the plurality of parent chromosome vector pairs to generate a plurality of (N+1)-th generation chromosome vectors, wherein N is 0 or a natural number, and the initial chromosome vector corresponds to a 0-th generation chromosome vector. A data processing system. A data processing system according to claim 7, wherein the policy decision circuit determines an optimal chromosome vector among the plurality of Nth generation chromosome vectors when N is greater than a predetermined maximum number of generations, stores the determined optimal chromosome vector in the profile database, and completes the optimization operation. A method of operating a data processing system comprising: a host executing a program for processing a data set; and a remote processing device connected to the host through an interface;
A profile generation step for storing a plurality of profiles corresponding to a plurality of functions included in the program while executing the program using a synthetic data set of a smaller size than the above data set; and
An optimization step for determining the execution location of a function corresponding to the plurality of profiles as either the host or the remote processing device while executing the program using the synthetic data set.
A method of operating a data processing system including: A method of operating a data processing system according to claim 9, wherein the profile generation step includes a step of generating a profile corresponding to a function from an array of return addresses of the function stored in stack memory when the function is called. A method of operating a data processing system according to claim 9, wherein the optimization step comprises the steps of generating an initial chromosome vector including a plurality of elements corresponding to the plurality of profiles, and arbitrarily setting a value corresponding to either the host or the remote processing device as a value of each element of the initial chromosome vector. A method of operating a data processing system according to claim 11, wherein the optimization step further comprises a step of adjusting a value of the initial chromosome vector so that the value has the same value as an element corresponding to a profile corresponding to an operation operation associated with the memory operation, if any one element of the initial chromosome vector is a profile corresponding to a memory operation. In claim 11, the optimization step
A step of generating a plurality of parent chromosome vector pairs according to a probabilistically selected order from a plurality of N-th generation chromosome vectors; and
A step of generating a plurality of (N+1) generation chromosome vectors by applying crossover and mutation techniques to the above plurality of parent chromosome vector pairs.
Including more than,
A method of operating a data processing system in which N is 0 or a natural number, and the initial chromosome vector corresponds to the 0th generation chromosome vector. In claim 13, the optimization step
A step of determining one of the Nth generation chromosome vectors as the optimal chromosome vector when the above N is greater than a pre-specified maximum number of generations.
A method of operating a data processing system further comprising: A method of operating a data processing system according to claim 14, further comprising the step of determining a plurality of execution locations corresponding to the plurality of profiles according to the optimal chromosome vector and executing the program for processing the data set, while moving a function called during execution of the program according to the plurality of execution locations to the host or the remote processing device.

Images