KR102644229B1 - Device and Method for Profiling Memory - Google Patents

Abstract

Disclosed is a memory profiling device and method of operating the same.
According to one aspect of the present invention, a memory profiling device for predicting access information of a memory object according to workload changes, comprising: an extraction unit for extracting access information of a memory object allocated by an application; an input unit for inputting the access information into a learned prediction model; an acquisition unit that obtains a scaling ratio of the memory object from the prediction model; and a prediction unit that predicts scaled access information of the memory object by applying the scaling ratio to the access information when the workload of the application changes.

Description

Memory profiling device and method of operation thereof {Device and Method for Profiling Memory}

Embodiments of the present invention relate to a machine learning-based memory profiling device and a method of operating the same that predict access information of a memory object according to workload changes through a prediction model in order to estimate the energy consumption of the memory object.

The content described in this section simply provides background information about the present invention and does not constitute prior art.

Recently, data processing volume has been increasing exponentially every year in various fields. To accommodate this data increase, companies and organizations are adopting optimized data centers. However, storing and analyzing data in data centers consumes a lot of energy and emits vast amounts of greenhouse gases.

Energy consumption in data centers is mainly determined by two factors: One is the Central Processing Unit (CPU), and the other is main memory. CPUs consume 30 to 60 percent of the total energy consumed by a single server, and main memory consumes 28 to 40 percent of the energy.

To reduce energy consumption at the memory-level, various technologies are being researched, including methods of lowering the power of memory banks, methods of controlling basic memory voltage and frequency, and hybrid memory systems that use both volatile and non-volatile memories. In addition, in memory systems, software-based solutions are being developed that additionally consider access information at the granular object-level to improve energy efficiency at the memory-level. In order to use this software-based solution, it is necessary to extract access information from memory objects and estimate energy consumption from the access information. Here, the process of extracting access information of memory objects is called memory profiling. A method for estimating energy consumption through memory profiling is disclosed in Korean Registration Notice No. KR 2178931 B1.

However, object-level memory profiling of an application has the problem that it takes a long time and consumes energy. In addition, when the application workload size changes, object access information also changes, so there is a problem that memory profiling must be performed every time the workload size changes.

Recently, the Linear Scaling Rate (LSR) model has been studied to reduce memory profiling time and predict access information of memory objects. The LSR model estimates that access information of memory objects changes linearly when the application workload changes. For example, if an application has a workload of size N and objects with a lifetime of L, it is assumed that when the workload size increases by a factor of two, the lifetime of the objects increases linearly by a factor of two. Through this, rather than extracting access information every time the size of the workload changes, the scaling ratio is applied to previous access information to predict access information corresponding to the changed workload.

However, since the LSR model assumes a linear relationship between workload and memory objects, it may contain prediction errors when applied to various applications. The smaller the correlation between the access information of a memory object and the scaling workload, the more error rate the access information of the memory object predicted based on the LSR model contains. In other words, when the application workload increases, the access information of memory objects does not increase linearly.

Therefore, when the size of the workload changes, access information is predicted using the scaling ratio to reduce time and energy consumption. Even if the workload and memory objects have various interrelationships, the access information of memory objects according to workload changes is Research is needed on how to accurately predict.

The main purpose of embodiments of the present invention is to provide a machine learning-based memory profiling device and a method of operating the same to reduce time and energy consumption for memory profiling according to workload changes.

The purpose of other embodiments of the present invention is to provide a memory profiling device and a method of operating the same for accurately predicting access information of memory objects according to workload changes, regardless of whether there is a linear relationship between the workload and the memory object. there is.

According to one aspect of the present invention, in a method of operating a memory profiling device that predicts access patterns of memory objects according to workload changes, access to memory objects allocated by an application The process of extracting information; A process of inputting the access information into a learned prediction model; Obtaining a scaling rate of the memory object from the prediction model; and predicting scaled access information of the memory object by applying the scaling ratio to the access information when the workload of the application changes.

According to another aspect of the present embodiment, a memory profiling device for predicting access information of a memory object according to a change in workload, comprising: an extraction unit for extracting access information of a memory object allocated by an application; an input unit that inputs the access information into a learned prediction model; an acquisition unit that obtains a scaling ratio of the memory object from the prediction model; and a prediction unit that predicts scaled access information of the memory object by applying the scaling ratio to the access information when the workload of the application changes.

As described above, according to an embodiment of the present invention, the scaling ratio of a memory object is predicted using a machine learning-based prediction model, and the scaling ratio is applied to the memory object access information even if the workload changes, thereby preventing workload changes. The time and energy consumption required to predict access information can be reduced.

According to another embodiment of the present invention, access information of memory objects that change according to workload changes can be accurately predicted using a machine learning-based prediction model.

Figure 1 is a diagram illustrating an application and a memory profiling device according to an embodiment of the present invention.
Figure 2 is a configuration diagram of a memory profiling device according to an embodiment of the present invention.
FIG. 3A is a diagram illustrating a memory profiling process according to an embodiment of the present invention.
Figure 3b is a diagram illustrating the process of learning a prediction model and estimating energy consumption using the prediction model according to an embodiment of the present invention.
Figure 4 is a diagram for explaining a process of extracting access information of a memory object according to an embodiment of the present invention.
Figures 5a and 5b are diagrams comparing memory profiling results of the present invention and the prior art.
Figure 6 is a flow chart illustrating a memory profiling process according to an embodiment of the present invention.

Hereinafter, some embodiments of the present invention will be described in detail through illustrative drawings. When adding reference numerals to components in each drawing, it should be noted that identical components are given the same reference numerals as much as possible even if they are shown in different drawings. Additionally, in describing the present invention, if it is determined that a detailed description of a related known configuration or function may obscure the gist of the present invention, the detailed description will be omitted.

Additionally, when describing the components of the present invention, terms such as first, second, A, B, (a), and (b) may be used. These terms are only used to distinguish the component from other components, and the nature, sequence, or order of the component is not limited by the term. Throughout the specification, when a part is said to 'include' or 'have' a certain component, this means that it does not exclude other components but may further include other components, unless specifically stated to the contrary. . Additionally, terms such as 'unit' and 'module' used in the specification refer to a unit that processes at least one function or operation, and may be implemented through hardware, software, or a combination of hardware and software.

Figure 1 is a diagram illustrating an application and a memory profiling device according to an embodiment of the present invention.

Referring to Figure 1, an application 100, a memory system 110, a memory profiler (112), and a memory device 114 are shown. Memory system 110 may include memory profiler 112 and memory device 114 .

The memory profiling device according to an embodiment of the present invention may be either the memory system 110 or the memory profiler 112.

The memory device 114 may be either a volatile memory or a non-volatile memory, or may be a heterogeneous memory device combining two types of memory. Here, the energy consumption of a memory object varies depending on the type of memory device 114. If the memory device 114 is a volatile memory, the energy information includes at least one of energy consumption consumed by preprocessing instructions, energy consumption consumed by read and write instructions, or energy consumption consumed by refresh instructions.

When the memory profiler 112 receives a profiling request from the application 100, it predicts the amount of energy consumed by a memory object allocated to the memory device 114 by the application 100.

This is because the type of the memory device 114 itself affects energy consumption, but the access pattern of the memory object used by the application 100 to the memory device 114 also affects energy consumption. For example, assume that there is a first memory device that consumes a lot of energy and a second memory device that consumes little energy when the memory writes data. At this time, when a memory object that performs a lot of writing is allocated to a first memory device, it consumes more energy than when it is allocated to a second memory device. In this way, the total energy consumption may vary depending on which memory object is used in which memory device.

Therefore, in order to accurately predict the energy consumption of a memory object, not only the characteristics of the memory device 114 itself but also the pattern in which the memory object accesses the memory device 114 must be considered. In particular, there is a need to extract access information that is highly relevant to energy consumption estimation.

When the memory profiler 112 according to an embodiment of the present invention receives a profiling request from the application 100, the memory object allocated to the memory device 114 by the application 100 accesses the memory device 114. Extract access information (access pattern). Here, the access information extracted for use in energy estimation includes at least one of the size of the memory object, the memory access size of the memory object (accessed volume), and the lifetime of the memory object. The memory profiler 112 may extract access information of a memory object by performing profiling one or more times.

However, when the memory profiler 112 receives a profiling request from the application 100, directly extracting access information of the memory object takes a lot of time and energy. In addition, when the size of the workload of the application 100 changes, the access information of the memory object also changes, so extracting the access information of the memory object every time the size of the workload changes is inefficient in terms of time and energy. am.

In contrast, the memory profiling device according to an embodiment of the present invention uses a prediction model to predict a scaling rate, which means the degree of change in access information according to workload changes, and when the workload changes By applying a scaling ratio to the memory object's access information, the memory object's access information for a changed workload can be predicted. And a new prediction model is used to differentiate it from the LSR model. Through this, the time and energy consumption required to extract access information of memory objects is reduced, and access information can be predicted more accurately, especially through a prediction model.

Figure 2 is a configuration diagram of a memory profiling device according to an embodiment of the present invention.

Referring to FIG. 2 , the memory profiling device 20 according to an embodiment of the present invention may include an extraction unit 200, an input unit 210, an acquisition unit 220, and a prediction unit 230. The memory profiling device 20 is a device for predicting access information of memory objects according to changes in the workload of an application.

The extraction unit 200 is a component that extracts access information of a memory object allocated by an application.

According to one embodiment of the present invention, the extraction unit 200 extracts access information related to the amount of energy consumed when a memory object is allocated to a memory device. Here, the access information used to estimate energy consumption includes at least one of the size of the memory object, the memory access size of the memory object, and the lifespan of the memory object.

As a specific extraction process, the extraction unit 200 can extract access information of a memory object in two steps. The first step is where the extraction unit 200 extracts the identifier of the memory object at the object level. The second step is a step in which the extraction unit 200 extracts at least one of the size, memory access size, and life time of the memory object at the instruction-level. More details are described in Figure 4.

The input unit 210 is a component that inputs access information into a learned prediction model.

The acquisition unit 220 is a component that acquires the scaling ratio of the memory object from the learned prediction model.

The prediction model is a model learned to receive access information of a memory object as input and output the scaling rate of the memory object.

The scaling ratio refers to the degree of change in access information according to workload changes. Specifically, the rate at which access information changes according to various workloads is called the scaling rate of access information of a memory object. The scaling ratio may include at least one of a scaling ratio for the size of the memory object, a scaling ratio for the memory access size, and a scaling ratio for the lifetime. Meanwhile, memory objects may have different scaling ratios for access information.

According to an embodiment of the present invention, the learned prediction model is a Linear Regression model, a K-Nearest Neighbor Regression model, and a Random Forest Regression model depending on the learning method. It can be any one of them.

The prediction unit 230 is a component that predicts scaled access information of a memory object by applying a scaling ratio to the access information when the workload of the application changes.

Accordingly, when the workload of an application changes, the memory profiling device 20 uses a scaling ratio obtained in advance through a prediction model rather than re-extracting access information according to the changed workload, thereby saving time and energy. Access information of memory objects can be predicted.

In addition, even if the access information of the workload and the memory object are in various relationships, including non-linear relationships, by using a prediction model, the scaling ratio and access information of the memory object according to the workload change can be accurately predicted. The types of prediction models are explained in detail in FIGS. 3A and 3B.

FIG. 3A is a diagram illustrating a memory profiling process according to an embodiment of the present invention.

Referring to FIG. 3A, when a memory profiling device receives a profiling request for a memory object from an application, it extracts access information of the memory object by performing a two-step memory profiling process. Here, the access information includes at least one of the size of the memory object, the memory access size of the memory object, and the lifespan of the memory object.

Figure 3b is a diagram illustrating the process of learning a prediction model and estimating energy consumption using the prediction model according to an embodiment of the present invention.

Referring to Figure 3b, the memory profiling device trains a prediction model based on the access information of the memory object, predicts the access information using the learned prediction model, and finally estimates the energy consumption of the memory object. They are shown in order.

Below, the specific learning process of the prediction model is explained.

The memory profiling device extracts access information about learning memory objects. Here, access information refers to information related to energy estimation.

Afterwards, the memory profiling device changes the workload and extracts access information of learning memory objects scaled according to the changed workload.

The memory profiling device can obtain scaling ratios of memory objects using the above information. Specifically, the memory profiling device calculates and obtains a scaling ratio for each memory object, which indicates the change in access information of the memory object according to the change in workload size.

Afterwards, the memory profiling device trains the prediction model on the access information and scaling ratio of the learning memory object. The prediction model learns the relationship between access information of memory objects and scaling ratio. The training data of the prediction model becomes the access information and scaling ratio of the memory object.

According to one embodiment of the present invention, when the prediction model is trained to receive access information of a memory object and output a scaling ratio, it also learns the correlation between the access information of the memory object. Here, the correlation between access information means whether a change in one access information affects other access information. For example, it means that there is a positive correlation between memory access size and lifespan among the access information of a memory object. In other words, the larger the memory access size of a memory object, the longer its lifespan. On the other hand, the size of the memory object is not correlated with other access information. The prediction model can learn not only the relationship between access information and scaling ratio of memory objects, but also the correlation between access information of memory objects.

Below, types of prediction models are described. The prediction model according to an embodiment of the present invention may be any one of a linear regression model, a K-nearest neighbor regression model, and a random forest regression model.

The linear regression model is a model that assumes that the application workload and access information of memory objects have a linear relationship, and that access information changes linearly as the workload changes. Scaling ratios between the workload and access information of learning memory objects are learned using a linear regression method.

According to an embodiment of the present invention, the linear regression model may apply linear regression for each memory object or may apply linear regression for each access information of the memory object.

The linear regression model predicts the scaling ratio using linear access information that exists between data sets, so when the access information has a linear relationship, the prediction accuracy of the scaling ratio is high. Additionally, in the linear regression model, the scaling ratio is predicted based only on linear access information, so the relationship between access information is not complicated and it has the characteristics of a light prediction model. However, if there is a correlation between the access information of the memory object of the workload, the linear regression model cannot accurately predict the scaling ratio. Therefore, the linear regression model is not suitable for predicting access information of various memory objects.

The K-nearest neighbor regression model is a model that predicts new data from the average of the K most recent data among existing data. The average of the scaling ratios of up to k learning memory objects is learned. The K-nearest neighbor regression model is a model trained to output the average of the scaling ratios of up to the kth memory objects when the scaling ratios of each memory object are different depending on the workload change.

The K-NNR model's prediction accuracy varies depending on the number of nearest neighbor data. If the k value is small, there may be overfitting to local characteristics of the data. Conversely, when the k value is large, the K-NNR model may be overgeneralized due to underfitting. From a prediction accuracy perspective, when there is a correlation between neighboring memory objects, the prediction accuracy according to the K-NNR model increases. The K-NNR model can prove the correlation between memory objects in an application and predict the scaling ratio. In addition, similar to the linear regression model, the K-NNR model predicts the scaling ratio using the k-th neighbors rather than predicting the scaling ratio from memory objects in complex relationships, so the K-NNR model is a lightweight model. You can.

The random forest regression model is a supervised machine learning model that finds patterns from the properties of each data and outputs the results. A tree is created by randomly extracting some of the learning data. Extracting part of the data is called bagging, and the result is predicted using a tree constructed by limiting the features of the data. That is, a tree is constructed using learning memory objects, and the scaling ratio of allocated memory objects is predicted using the tree.

According to an embodiment of the present invention, when the prediction model is a random forest regression (RFR) model, the number of trees constituting the RFR must be considered for the accuracy of the prediction model. Since the accuracy of the prediction model varies depending on the number of trees, each application finds the point where accuracy saturates depending on the number of trees and reflects it in the learning process. To this end, the number of trees of the learned prediction model may be determined based on the scaling ratios of memory objects for learning. Specifically, the number of trees of the learned prediction model may be determined to be higher as the scaling ratios of learning memory objects are more irregular. This is to increase the prediction accuracy of the scaling ratio and access information by increasing the learning amount of the prediction model by increasing the number of trees as the relationship between the workload and the access information of the memory object becomes more complex.

When the trained prediction model receives access information of a memory object, it outputs the scaling ratio of the memory object.

Figure 4 is a diagram for explaining a process of extracting access information of a memory object according to an embodiment of the present invention.

Referring to Figure 4, the memory profiling device extracts access information of a learning memory object through two stages including a fast stage and a slow stage. This is called Two Pass Memory Profiling (TPMP).

As a first step, the memory profiling device performs primary profiling on the learning memory object. Primary profiling is the first step and corresponds to the quick step. In a quick step, the memory profiling device extracts the hash value of the function call stack to distinguish different variables in the application and map the objects of each variable to the corresponding variable. Here, the hash value corresponds to the identifier of the memory object.

Specifically, the memory profiling device intercepts dynamic allocation functions such as malloc, calloc, and realloc functions using a wrapper library, then stores the call stack of the dynamic allocation function as a string and hashes it. . The memory profiling device extracts the hash values of all memory objects and writes them to the result file. Then, offline processing is performed, key variables for secondary profiling are selected, and the hash values are listed in the file.

Afterwards, the memory profiling device performs secondary profiling on the learning memory object. Secondary profiling is the second stage and corresponds to the slow stage. In the slow stage, the memory profiling device performs profiling of access information on key variables at the instruction level using a PIN tool. At this time, through the pin tool library, the memory profiling device can extract only the access information necessary for energy estimation and omit access information unrelated to energy estimation.

For example, the memory profiling device can use the Intel Pin tool to profile access to all learning memory objects at runtime. In this case, the profiling results for learning memory objects are profiled in the first step. You can refer to the created hash table and save the result of the learning memory object to which the object belongs.

As a second step, the memory profiling device performs secondary profiling on the learning memory object to extract at least one of the size, memory access size, and lifespan of the learning memory object.

The memory profiling device uses the extracted access information to determine the predicted energy consumption when a learning memory object is allocated to a memory device. Specifically, the memory profiling device determines energy information consumed according to instructions performed on a memory device due to a learning memory object, and determines energy consumption using access information and energy information.

Meanwhile, in this specification, the memory profiling device is described as using two-pass memory profiling that profiles through the first and second steps to extract access information of the learning memory object, but access to the learning memory object You can also use other memory profiling techniques to extract information.

Figures 5a and 5b are diagrams comparing memory profiling results of the present invention and the prior art.

Referring to FIG. 5A, a graph comparing the energy estimation accuracy by application between the linear scaling ratio model of the prior art and the memory profiling device of the present invention is shown. The graph shows that the prediction accuracy of energy consumption estimation according to the memory profiling device is higher than that of the prior art in several applications.

Referring to FIG. 5B, a graph is shown comparing the accuracy of predicting access information of a memory object by application between the linear scaling ratio model of the prior art and the memory profiling device of the present invention. The memory profiling device predicts access information through a scaling ratio obtained from a prediction model, and the prior art predicts access information through a linear scaling ratio.

When the prediction model is an RFR model, the access information prediction accuracy of the memory profiling device is 92.85 percent on average, and the energy consumption estimation accuracy is 91.3 percent.

Figure 6 is a flowchart illustrating a memory profiling process according to an embodiment of the present invention.

Referring to FIG. 6, the memory profiling device extracts access information of the memory object allocated by the application (S600). Here, the access information includes at least one of the size, accessed volume, and lifetime of the memory object.

The specific extraction process is performed in two steps, including extracting the identifier of the memory object at the object level and extracting at least one of the size, memory access size, and lifetime time of the memory object at the instruction level.

The memory profiling device inputs access information into the learned prediction model (S602). Here, the prediction model is trained to receive access information of learning memory objects and output an output scaling ratio. Depending on the type of learning, the prediction model may be one of a Linear Regression model, K-Nearest Neighbor Regression model, and Random Forest Regression model. If the learned prediction model is a random forest regression model, the number of trees of the learned prediction model may be determined based on the scaling ratios of memory objects for learning. Specifically, the more irregular the scaling ratios of learning memory objects are, the higher the number of trees of the learned prediction model can be determined.

The memory profiling device obtains the scaling ratio of the memory object from the prediction model (S604).

When the workload of an application changes, the memory profiling device predicts the scaled access information of the memory object by applying the scaling ratio to the access information (S606).

Later, the memory profiling device can estimate the energy consumption of the memory object from the scaled access information obtained using the scaling ratio.

In Figure 6, processes S600 to S606 are described as being sequentially executed, but this is merely an exemplary explanation of the technical idea of an embodiment of the present invention. In other words, a person skilled in the art to which an embodiment of the present invention pertains can change the sequence shown in FIG. 6 and perform one of steps S600 to S606 without departing from the essential characteristics of the embodiment of the present invention. Since the above processes can be applied in various modifications and variations by executing them in parallel, FIG. 6 is not limited to a time series order.

Meanwhile, the processes shown in FIG. 6 can be implemented as computer-readable codes on a computer-readable recording medium. Computer-readable recording media include all types of recording devices that store data that can be read by a computer system. In other words, such computer-readable recording media include non-transitory media such as ROM, RAM, CD-ROM, magnetic tape, floppy disk, and optical data storage devices. Additionally, computer-readable recording media can be distributed across networked computer systems so that computer-readable code can be stored and executed in a distributed manner.

Additionally, the components of the present invention may use direct circuit structures such as memory, processor, logic circuit, look-up table, etc. These integrated circuit structures execute each function described herein through control of one or more microprocessors or other control devices. Additionally, the components of the present invention may be specifically implemented by a program or portion of code that includes one or more executable instructions for performing specific logical functions and is executed by one or more microprocessors or other control devices. Additionally, the components of the present invention may include or be implemented by a central processing unit (CPU), a microprocessor, etc. that perform their respective functions. Additionally, the components of the present invention may store instructions executed by one or more processors in one or more memories.

The above description is merely an illustrative explanation of the technical idea of the present embodiment, and those skilled in the art will be able to make various modifications and variations without departing from the essential characteristics of the present embodiment. Accordingly, the present embodiments are not intended to limit the technical idea of the present embodiment, but rather to explain it, and the scope of the technical idea of the present embodiment is not limited by these examples. The scope of protection of this embodiment should be interpreted in accordance with the claims below, and all technical ideas within the equivalent scope should be interpreted as being included in the scope of rights of this embodiment.

20: memory profiling device 200: extractor
210: input unit 220: acquisition unit
230: prediction unit

Claims (14)

In a method of operating a memory profiling device that predicts access patterns of memory objects according to workload changes,
A process of extracting access information of a memory object allocated by an application;
A process of inputting the access information into a learned prediction model;
Obtaining a scaling rate of the memory object from the prediction model; and
When the workload of the application changes, the process of predicting scaled access information of the memory object by applying the scaling ratio to the access information.
A method of operating a memory profiling device comprising: According to paragraph 1,
The above access information is,
A method of operating a memory profiling device including at least one of size, accessed volume, and lifetime of the memory object. According to paragraph 1,
The extraction process is,
A process of extracting the identifier of the memory object at the object-level; and
A process of extracting at least one of the size, memory access size, and lifetime time of the memory object at the instruction-level
A method of operating a memory profiling device comprising: According to paragraph 1,
The learned prediction model is,
A method of operating a memory profiling device that is trained to receive access information of learning memory objects and output an output scaling ratio. According to paragraph 1,
The learned prediction model is,
A method of operating a memory profiling device that is one of a Linear Regression model, a K-Nearest Neighbor Regression model, and a Random Forest Regression model. According to paragraph 1,
When the learned prediction model is a random forest regression model, the number of trees of the learned prediction model is determined based on scaling ratios of memory objects for learning. According to clause 6,
The number of trees of the learned prediction model is,
A method of operating a memory profiling device, wherein the more irregular the scaling ratios of the learning memory objects are, the higher they are determined. In a memory profiling device that predicts access information of memory objects according to workload changes,
an extraction unit that extracts access information of a memory object allocated by an application;
an input unit for inputting the access information into a learned prediction model;
an acquisition unit that obtains a scaling ratio of the memory object from the prediction model; and
When the workload of the application changes, a prediction unit that predicts the scaled access information of the memory object by applying the scaling ratio to the access information.
A memory profiling device including a. According to clause 8,
The above access information is,
A memory profiling device including at least one of a size, memory access size, and lifetime of the memory object. According to clause 8,
The extraction unit,
A memory profiling device that extracts the identifier of the memory object at the object-level and extracts at least one of the size, memory access size, and lifetime time of the memory object at the instruction-level. According to clause 8,
The learned prediction model is,
A memory profiling device that is trained to receive access information of learning memory objects and output an output scaling ratio. According to clause 8,
The learned prediction model is,
A memory profiling device that is either a linear regression model, a K-nearest neighbor regression model, or a random forest regression model. According to clause 8,
When the learned prediction model is a random forest regression model, the number of trees of the learned prediction model is determined based on scaling ratios of memory objects for learning. According to clause 13,
The number of trees of the learned prediction model is,
A memory profiling device in which the scaling ratios of the learning memory objects are determined to be higher as they become more irregular.

Images