KR102469927B1 - Apparatus for managing disaggregated memory and method for the same - Google Patents

Abstract

A divided memory management method is disclosed. A divided memory management method according to the present disclosure includes allocating at least one memory page to a local memory and a remote memory, checking a request for access to the memory page, and the memory that is requested to be accessed. A process of checking whether a target performance ratio required by a service is satisfied as the page is allocated to the remote memory; It may include a process of estimating the size of , and a process of reallocating the memory page requested for access in consideration of the predicted size of the local memory.

Description

Split memory management device and method {APPARATUS FOR MANAGING DISAGGREGATED MEMORY AND METHOD FOR THE SAME}

The present disclosure relates to a computing system, and more particularly, to a method and apparatus for controlling the size of memory allocated to a virtual machine.

With the development of networking technology, a cloud system for sharing computing resources is being provided. A cloud system may be configured so that multiple users can share various physical infrastructures. The physical infrastructures may include one or more computing systems having processors, memory, storage, networking, and the like.

Physical infrastructure must be used to implement or run the workloads required by the application. However, since the physical infrastructure must be shared within the cloud system, the cloud system manages the physical infrastructure to implement or execute the workload required by the application through a logical server or virtual machine (VM). are doing

Furthermore, due to the rapid increase in data-intensive workloads such as in-memory databases, data caching, bioinformatics, and graph processing, the demand for cloud systems Memory capacity is increasing.

In line with this trend, a large-capacity cloud system capable of using 1 TB or more of memory in a single virtual machine (VM) is being provided.

As described above, a large-capacity physical server infrastructure is required to provide a large-capacity type cloud system, but there is a problem in that a lot of cost is consumed to build such an infrastructure.

In order to support large amounts of memory by using multiple small-volume general-purpose servers instead of large-capacity physical servers, a disaggregated memory structure technology that operates memory distributed in different physical computing devices (e.g., servers) as a single memory will be used. can

An object of the present disclosure is to provide an apparatus and method capable of efficiently managing a memory in a divided memory structure.

Another technical problem of the present disclosure is to provide an apparatus and method capable of managing memory while guaranteeing contract performance required in a cloud system configured with a divided memory structure.

Another technical problem of the present disclosure is to provide an apparatus and method for providing a memory that can guarantee required contract performance through LRU distance-based performance prediction.

The technical problems to be achieved in the present disclosure are not limited to the technical problems mentioned above, and other technical problems not mentioned will be clearly understood by those skilled in the art from the description below. You will be able to.

According to one aspect of the present disclosure, a divided memory management method may be provided. In the memory management method, the method includes allocating at least one memory page to a local memory and a remote memory, confirming a request for access to the memory page, and allocating the requested memory page to a remote memory. process of checking whether the target performance ratio required by the service is satisfied, and predicting the size of the local memory based on the histogram based on the LRU distance in response to not satisfying the target performance ratio and reallocating the memory page requested for access in consideration of the predicted size of the local memory.

According to another aspect of the present disclosure, a partitioned memory management device may be provided. The device includes a divided memory unit including a local memory provided in a computing device in a local area and a remote memory provided in a computing device in a remote area, and a memory managing at least one memory page required for building or executing an application program. A virtual machine unit, a memory hypervisor unit provided between the divided memory unit and the memory virtual machine unit, and allocating the at least one memory page to a local memory and a remote memory, and confirming a request for access to the memory page; As the memory page requested for access is allocated to the remote memory, it is checked whether the target performance ratio required by the service is satisfied, and in response to the target performance ratio not being satisfied, based on the histogram based on the LRU distance A performance predictor for estimating the size of the local memory may be included.

According to another aspect of the present disclosure, a partitioned memory management device may be provided. The device includes a divided memory unit including a local memory provided in a computing device in a local area and a remote memory provided in a computing device in a remote area, and a memory managing at least one memory page required for building or executing an application program. A virtual machine unit, provided between the divided memory unit and the memory virtual machine unit, allocates the at least one memory page to a local memory and a remote memory, checks a request for access to the memory page, and requests access to the memory As the page is allocated to the remote memory, it is checked whether the target performance ratio required by the service is satisfied, and in response to the target performance ratio not being satisfied, the size of the local memory based on the histogram based on the LRU distance. and a memory hypervisor unit for reallocating the memory page requested for access in consideration of the predicted size of the local memory.

The features briefly summarized above with respect to the disclosure are merely exemplary aspects of the detailed description of the disclosure that follows, and do not limit the scope of the disclosure.

According to the present disclosure, an apparatus and method capable of efficiently managing a memory in a divided memory structure may be provided.

In addition, according to the present disclosure, it is possible to manage memory while guaranteeing contract performance required by a cloud system configured with a divided memory structure.

In addition, according to the present disclosure, it is possible to provide a memory capable of guaranteeing a required contract performance through performance prediction based on an LRU distance.

Effects obtainable in the present disclosure are not limited to the effects mentioned above, and other effects not mentioned may be clearly understood by those skilled in the art from the description below. will be.

1 is a diagram illustrating a system to which a partitioned memory management device according to an embodiment of the present disclosure is applied.
2 is a block diagram illustrating the structure of a partitioned memory management device according to an embodiment of the present disclosure.
3 is a diagram illustrating an LRU distance histogram generated in a partitioned memory management device according to an embodiment of the present disclosure.
4 is a flowchart illustrating a sequence of a divided memory management method according to an embodiment of the present disclosure.
5 is a flowchart illustrating a detailed sequence of a process of estimating memory performance based on an LRU distance included in a partitioned memory management method according to an embodiment of the present disclosure.

Hereinafter, with reference to the accompanying drawings, embodiments of the present disclosure will be described in detail so that those skilled in the art can easily carry out the present disclosure. However, the present disclosure may be implemented in many different forms and is not limited to the embodiments described herein.

In describing the embodiments of the present disclosure, if it is determined that a detailed description of a known configuration or function may obscure the gist of the present disclosure, a detailed description thereof will be omitted. And, in the drawings, parts irrelevant to the description of the present disclosure are omitted, and similar reference numerals are attached to similar parts.

In the present disclosure, when a component is said to be "connected", "coupled" or "connected" to another component, this is not only a direct connection relationship, but also an indirect connection relationship between which another component exists. may also be included. In addition, when a component "includes" or "has" another component, this means that it may further include another component without excluding other components unless otherwise stated. .

In the present disclosure, terms such as voice parameter and second are used only for the purpose of distinguishing one component from another, and do not limit the order or importance of components unless otherwise specified. Accordingly, within the scope of the present disclosure, a voice parameter component in one embodiment may be referred to as a second component in another embodiment, and similarly, a second component in one embodiment may be referred to as a voice parameter component in another embodiment. can also be called

In the present disclosure, components that are distinguished from each other are intended to clearly explain each characteristic, and do not necessarily mean that the components are separated. That is, a plurality of components may be integrated to form a single hardware or software unit, or a single component may be distributed to form a plurality of hardware or software units. Accordingly, even such integrated or distributed embodiments are included in the scope of the present disclosure, even if not mentioned separately.

In the present disclosure, components described in various embodiments do not necessarily mean essential components, and some may be optional components. Therefore, an embodiment composed of a subset of components described in one embodiment is also included in the scope of the present disclosure. In addition, embodiments including other components in addition to the components described in various embodiments are also included in the scope of the present disclosure.

Hereinafter, embodiments of the present disclosure will be described with reference to the accompanying drawings.

1 is a diagram illustrating a system to which a partitioned memory management device according to an embodiment of the present disclosure is applied.

Referring to FIG. 1 , a system 100 to which a partitioned memory management apparatus is applied includes a resource 110 , a hypervisor 120 and a virtual machine 130 .

The resource 110 may include a plurality of racks 112-1, 112-2, ≤. 112-n. Each of the racks 112-1, 112-2, ≤. 112-n may include various computing resources. These computing resources may include several types of physical resources. For example, physical resources include CPU, memory (eg, random access memory (RAM)), storage (eg, hard disk or solid state drive), and NW I/O devices (eg, network interfaces). cards), power devices (eg, power bricks), cooling devices (eg, fans or coolants), or other types of resources (eg, network switch devices).

These computing resources are available through the hypervisor 120 (eg, to a resource manager or controller). The hypervisor 120 may be assigned to the virtual machine 130 by a resource manager, controller, or scheduler of the system 100 . In particular, the hypervisor 120 may calculate scores for a plurality of computing resources and evaluate a rank for each computing resource. Ranking of computing resources may be performed for elements that can meet power, performance, cost, availability, or maintenance requirements.

The hypervisor 120 may monitor a plurality of operating attributes for each computing resource in order to configure or deploy the virtual machine 130 while the virtual machine 130 implements, runs, or executes a workload.

In addition, the hypervisor 120 may allocate computing resources to the virtual machine 130 based on scores of the computing resources. For example, hypervisor 120 may realize a required load from a previous task or rank that operating score compared to past operating scores determined for one or more other portions of the computing resource. Specifically, when the ranking of the first computing resource is evaluated lower than the previous operation scores of other computing resources, the hypervisor 120 may replace the first computing resource with a new computing resource.

Meanwhile, the virtual machine 130 may include a plurality of logical servers or virtual machines 132-1, 132-2, ≤. 132-m. For example, logical servers or virtual machines 132-1, 132-2, ≤. 132-m may be configured to implement or execute workloads. Each logical server required for the virtual machine 130 may include one or more virtual machines. In addition, a portion of computing resources may be allocated to each virtual machine. As another example, computing resources may be directly allocated to specific virtual machines.

Furthermore, the memory included in the computing resource may be provided in a disaggregated memory structure. Here, the disaggregated memory structure may include a plurality of memories physically distributed in different devices (eg, server devices). Also, the plurality of memories can be divided into a local memory located in the same device as the computing device and a remote memory located in a device remote from the computing device based on the location where the memory is located.

Such a divided memory structure can solve the problem of memory usage imbalance by dynamically providing memory that is not used by a device to other devices, and at the same time, it is possible to increase a memory usage rate. Furthermore, since a plurality of devices share and use the memory, a large amount of memory can be used even if the plurality of devices do not have a large amount of memory that can be used locally.

2 is a block diagram illustrating the structure of a partitioned memory management device according to an embodiment of the present disclosure.

An apparatus for managing partitioned memory according to an embodiment of the present disclosure may be configured based on the system 100 of FIG. 1 described above.

As described above, physical resources include CPU, memory (eg, random access memory (RAM)), storage (eg, hard disk or solid state drive), and NW I/O devices (eg, network interface cards), power devices (e.g. power bricks), cooling devices (e.g. fans or coolants), or other types of resources (e.g. network switch devices), etc. However, the divided memory management device is a device provided to manage memory among various physical resources, and may include configurations required for memory management.

Referring to FIG. 2 , the partitioned memory management apparatus 200 may include a partitioned memory unit 210 , a memory hypervisor unit 220 , and a memory virtual machine unit 230 .

The divided memory unit 210 includes a local memory 211 located in the same device 200 as the device where computing is performed, and a remote memory 215 located in a device 200' remote from the device where computing is performed. can do.

The memory hypervisor unit 220 performs memory management so that the memory virtual machine unit 230 can recognize the local memory 211 and the remote memory 215 provided in the divided memory unit 210 as a single memory. can

The memory hypervisor unit 220 may remove the complexity of accessing the remote remote memory 215 through remote memory access in page units. Specifically, the memory hypervisor unit 220 generates a page fault when the memory page requested to be accessed by the memory virtual machine unit 230 does not exist in the local memory 211, and through page fault processing. The remote memory 215 may be fetched as a local page. The memory hypervisor unit 220 may allocate and manage the fetched local page to the memory virtual machine unit 230 where a page fault has occurred. Accordingly, when access to the corresponding memory is requested, the memory virtual machine unit 230 may access and process the fetched local page.

As another example, when a page available in the local memory 211 does not exist, the memory hypervisor unit 220 may evict the local memory page to the remote memory 215, and through this operation, the local memory page A memory page can be freed.

Furthermore, the memory hypervisor unit 220 monitors the size of the memory used by the memory virtual machine unit 230 in order to smoothly provide the memory of the capacity required by the memory virtual machine unit 230, and the memory virtual machine It is necessary to estimate the size of memory required in unit 230.

Hereinafter, an operation of predicting the memory size by the memory hypervisor unit 220 will be described in detail.

An operating system or an application program may be executed through the memory virtual machine unit 230. When the application program is executed, the execution time of the application program may increase in proportion to the remote memory 215 access number n.

Based on this, the application program execution time (T(n)) for the remote memory 215 access number (n) can be expressed as in Equation 1 below.

Here, L is the execution time of an application in the memory virtual machine unit 230 when the local memory 211 is sufficient, A is the average time required to access the remote memory 215, and n is the access to the remote memory 215 represents a number of times.

The application program execution time (L) is a fixed value independent of n, and the value can be checked after the application program is executed. The application execution time (L) is determined by Equation 2 and can be shown together.

Meanwhile, the performance degradation rate can be expressed by Equation 3 below.

Since the application program execution time (L) can be known using Equation 2 described above, by substituting this into Equation 1, the application execution time according to n can be calculated as in Equation 4 below.

Meanwhile, when providing a service (eg, cloud service), it is necessary to guarantee a performance ratio (hereinafter referred to as 'target performance ratio') required by the service (eg, cloud service), so the memory hypervisor unit (220 ) needs to consider the target performance ratio when estimating the size of the memory.

The target performance ratio α represents a performance ratio guaranteed by the service and may be set to a value between 0 and 1. For example, when the value of the target performance ratio (α) is 0.9, compared to the case where the entire contracted memory in the service is allocated to the memory virtual machine unit 230, the application program executed in the memory virtual machine unit 230 It means that the performance is guaranteed at least 90% or more.

Based on this, the memory hypervisor unit 220 needs to measure the performance of the application program in order to guarantee the contracted performance of these services, and predict the size of memory to be additionally input to compensate for the insufficient performance.

To guarantee the target performance ratio (α), the execution time of the application program must be relatively less than or equal to L/α. In order to reduce the execution time of the application program by PL/α, the number of remote memory accesses must be reduced by at least Nn _t according to the relationship represented by Equation 5 below.

And, referring to Equations 4 and 5 above, the number of remote memory accesses (Nn _t ) to be reduced can be expressed as Equation 6 below.

Meanwhile, the memory hypervisor unit 220 may calculate the size of the local memory 211 additionally required to reduce the number of remote memory accesses by Nn _t .

At this time, the memory hypervisor unit 220 may calculate the additionally required size of the local memory 211 by utilizing a page LRU distance histogram for the remote memory 215 .

Specifically, the memory hypervisor unit 220 may allocate frequently used memory pages among memory pages allocated to the memory virtual machine unit 230 to the local memory 211, and infrequently used memory pages to the remote memory. (215) can be assigned.

Also, the memory hypervisor unit 220 may use the local memory 211 as a memory cache. In addition, the memory hypervisor unit 220 may calculate the LRU distance of an ejected page to the remote memory 215 and generate an LRU distance histogram based on the calculated LRU distance.

For example, if the number of memory pages allocated to the memory virtual machine unit 230 is 8, the memory hypervisor unit 220 allocates 4 pages to the remote memory 215 and 4 pages to the local memory 211. pages can be allocated.

Specifically, when the allocated memory pages are A, B, C, D, E, F, G, and H pages, the memory hypervisor unit 220 stores A, B, C, and D pages in the local memory 211. and allocate E, F, G, and H pages to the remote memory 215.

As shown in Table 1, when the page access order of the memory virtual machine unit 230 is E, A, D, F, C, H, the LRU distance to the remote memory 215 can be calculated as shown in Table 1 below. . In Table 1, the left side of the character string representing the page represents the MRU page in the LRU list, and the right side of the character string represents the LRU page. The LRU distance value indicates the number of accesses to each location of the remote memory 211 page.

The memory hypervisor unit 220 may generate the LRU distance histogram shown in FIG. 3 by cumulatively calculating the LRU distance values of Table 1 described above, and this LRU distance histogram may be expressed in a function form.

Furthermore, the LRU distance histogram function H(c) may correspond to the number of accesses to the remote memory 215 based on the LRU distance.

Since access to the remote memory 215 may be caused by a shortage of the local memory 211 , the number of accesses to the remote memory 215 may correspond to the shortage of the local memory 211 . In other words, when the local memory 211 is allocated a memory size c equal to the number of accesses to the remote memory 215 based on the LRU distance, access to the remote memory 215 may not occur. Therefore, the LRU distance histogram function (H(c)) and the number of remote memory accesses to be reduced (Nn _t ) are set to be the same, and from this, the shortfall of the local memory 211, that is, the local memory 211 to be added It can be set to the size (c) of Based on this, the memory hypervisor unit 220 may calculate the size (c) of the local memory 211 to be added through the operation of Equation 7 below.

The LRU distance histogram function (H(c)), the average time required to access the remote memory 215 (A), and the number of accesses to the remote memory 215 (N) can be obtained by profiling the execution of the application. Therefore, the memory hypervisor unit 220 continuously monitors the performance degradation within a predetermined time period to predict performance, and based on this, the insufficient amount of the local memory 211, that is, the amount of the local memory 211 to be added. size (c) can be predicted.

As described above, the memory hypervisor unit 220 exemplifies performing an operation of predicting the size c of the local memory 211 to be added based on Equations 1 to 7, but the present disclosure This is not limiting. As another example, the partitioned memory management apparatus 200 may include a performance predictor 250 configured separately from the memory hypervisor unit 220, and the performance predictor 250 is based on Equations 1 to 7 above. , an operation of estimating the size (c) of the local memory 211 to be added may be performed.

Meanwhile, the memory hypervisor unit 220 checks the size (c) of the local memory 211 to be added, and the remaining memory of the local memory 211 is equal to or equal to the size (c) of the local memory 211 to be added. When it has a relatively large value, the memory page may be allocated to the local memory 211 . On the other hand, if the remaining memory of the local memory 211 has a relatively smaller value than the local memory 211 to be added, the memory hypervisor unit 220 will migrate the virtual machine to another node with sufficient local memory. can

Meanwhile, since the memory virtual machine unit 230 transfers the insufficient local memory size to the memory hypervisor unit 220, the size of the local memory page allocated to the memory virtual machine unit 230 is the working set size of the application ( working set size), the partitioned memory management device may not recognize it. In this case, since the number of memory pages used by the real memory virtual machine unit 230 may be relatively smaller than the memory pages allocated to the memory virtual machine unit 230, the local memory usage efficiency may be lowered.

In one embodiment of the present disclosure, the memory hypervisor unit 220 has a predetermined size (γ, 0 < The size of the page allocated to the local memory can be reduced by γ≤ 1).

In this way, even if the size of the page allocated to the local memory is reduced by a predetermined size (γ), the memory page is reallocated by periodically checking the size of the local memory every predetermined time unit, so the memory virtual machine unit 230 ) can secure as much local memory as required.

Accordingly, the apparatus for managing partitioned memory can increase memory usage efficiency by adaptively allocating memory pages according to the use state of resources, and can stably allocate memory pages by estimating the size of a local memory.

4 is a flowchart illustrating a sequence of a divided memory management method according to an embodiment of the present disclosure.

The divided memory management method may be performed by the above-described divided memory management device.

First, the partitioned memory management device may request access to a memory page by a virtual machine (S401).

The memory included in the divided memory management apparatus may be provided in a disaggregated memory structure. Here, the disaggregated memory structure may include a plurality of memories physically distributed in different devices (eg, server devices). Also, the plurality of memories can be divided into a local memory located in the same device as the computing device and a remote memory located in a device remote from the computing device based on the location where the memory is located.

Based on this, if the memory page to which access is requested is a remote memory (S402-Yes), the divided memory management device may bring the remote memory page to the local memory page (S403). And, the partitioned memory management apparatus may create and manage remote memory access information notifying that remote memory access has occurred (S404). Here, the remote memory access information may include a remote memory identifier, a time when a remote memory page is fetched, and the number of remote memory accesses.

Next, the partitioned memory management apparatus may perform memory performance prediction based on the LRU distance (S405).

Memory performance prediction based on the LRU distance is an operation of estimating the size (c) of the local memory 211 to be added based on Equations 1 to 7, and is performed by a memory hypervisor provided in the divided memory management device. It can be. As another example, memory performance prediction based on the LRU distance may be performed by a performance predictor provided separately from the memory hypervisor.

Memory performance estimation based on the LRU distance will be described in detail in an operation related to FIG. 5 to be described later.

Meanwhile, when the memory page to which access is requested is a local memory (S402-No), the divided memory management apparatus may proceed to step S406.

In step S406, the partitioned memory management apparatus may check the state of the local memory 211 to determine whether the remaining memory of the local memory 211 is sufficiently secured as the size of the memory to which access is requested. For example, the partitioned memory management apparatus may check whether the remaining memory size of the local memory 211 is equal to or relatively larger than the size of the memory to which access is requested.

If the remaining memory size of the local memory 211 is equal to or relatively large than the size of the memory for which access is requested (S407-Yes), the partitioned memory management device allocates the memory page for which access is requested to the local memory. It can (S408). On the other hand, if the size of the remaining memory of the local memory 211 indicates a relatively smaller value than the memory size for which access is requested (S407-No), the virtual machine can be migrated to another node with sufficient local memory. (S409).

5 is a flowchart illustrating a detailed sequence of a process of estimating memory performance based on an LRU distance included in a partitioned memory management method according to an embodiment of the present disclosure.

First of all, when providing a service (eg, cloud service), it is necessary to guarantee the performance ratio (hereinafter referred to as 'target performance ratio') required by the service (eg, cloud service). When estimating the size of , it is necessary to consider the target performance ratio. Based on this, in step S501, the divided memory management device may set a target performance ratio α. The partitioned memory management device may provide an environment (e.g., a user interface) for inputting the target performance ratio (α), receive the target performance ratio (α) using this environment, and determine the target performance ratio (α). can be set.

Next, in step S502, the partitioned memory management device sets parameter values (start time (S), memory monitoring time period (P), LRU distance histogram function (H), remote memory 215 to be used for predicting memory performance based on LRU distance). ), the average time (A) required to access the remote memory 215, the number of accesses (N), etc.) may be initialized.

Thereafter, access to the local memory or remote memory may be requested by the virtual machine. When the remote memory access request occurs (S503-Yes), remote memory access information notifying that the remote memory access has occurred is generated and can be provided. Accordingly, the partitioned memory management device can check the remote memory access information (S504). Here, the remote memory access information may include a remote memory identifier, a time when a remote memory page is fetched, and the number of remote memory accesses.

Thereafter, the divided memory management device may update the remote memory access count by increasing 1 (S505). Then, the partitioned memory management apparatus may check the remote memory page to which access is requested and the LRU distance of the remote memory page to which access is requested (S506).

Next, the partitioned memory management device removes the access page from the LRU list of the remote memory (S507), and if there is an ejected page from the local memory to the remote memory, it can be added to the MRU location of the LRU list of the remote memory. Yes (S508).

Then, the divided memory management device considers the relationship between the application execution time (L), the monitoring time period (P), the number of remote memory accesses (N), and the average remote memory access time (A) described in Equation 2 above. Thus, the average remote memory access time (A) can be calculated (S509).

Then, the divided memory management device may compare the current time with the start time (S) and the monitoring time period (P). If the current time represents a value equal to or relatively greater than the sum of the start time (S) and the monitoring time period (P) (S510-Yes), the divided memory management device proceeds to step S511, and the current time is the start time ( S) and the monitoring time period (P) when the value is relatively smaller than the sum (S510-No), step S503 may be performed.

In step S511, the divided memory management device may check the application program execution time L, and calculate the application program execution time T(n) according to n based on the description in Equation 4 above.

In step S512, the divided memory management device determines that the application execution time (L) reflecting the target performance ratio (α) (ie, the application execution time (L)/target performance ratio (α)) corresponds to the monitoring time period (P). You can check whether they represent the same or relatively large values.

If the application execution time (L) reflecting the target performance ratio (α) has the same or relatively large value as the monitoring time period (P), it means that the target performance is satisfied, so the partitioned memory management device is a parameter value to be used for performance prediction. A step S502 of performing initialization of s may be performed.

On the other hand, if the application execution time (L) reflecting the target performance ratio (α) is relatively smaller than the monitoring time period (P), it means that the target performance is not satisfied, so the divided memory management device performs step S513. can proceed

In step S513, the divided memory management device may generate a histogram function H(c) based on the LRU distance value checked in step S506. And, in step S514, the divided memory management device calculates the size (c) of the local memory to be added based on the relationship between the above-described LRU distance histogram function (H(c)) and the number of remote memory accesses (Nn _t ) to be reduced and can provide.

On the other hand, since the partitioned memory management device transmits the size of the local memory lacking in the virtual machine to the hypervisor, even if the size of the local memory page allocated to the virtual machine is larger than the working set size of the application, the partitioned memory The management device may not be aware of it. In this case, since the memory pages actually used by the virtual machine may be relatively smaller than the memory pages allocated to the virtual machine, the local memory usage efficiency may be lowered.

In one embodiment of the present disclosure, the partitioned memory management device is configured to store data in local memory by a predetermined size (γ, 0 <γ≤ 1) when a virtual machine does not access a remote memory for a certain period of time or when the vcpu usage rate of the virtual machine is low. You can reduce the size of allocated pages.

In this way, even if the size of the page allocated to the local memory is reduced by a predetermined size (γ), the size of the local memory is periodically checked every predetermined time unit and the memory page is reallocated. of local memory can be secured.

Accordingly, the divided memory management method can increase memory usage efficiency by adaptively allocating memory pages according to the use state of resources, and can stably allocate memory pages by estimating the size of a local memory.

Exemplary methods of this disclosure are presented as a series of operations for clarity of explanation, but this is not intended to limit the order in which steps are performed, and each step may be performed concurrently or in a different order, if desired. In order to implement the method according to the present disclosure, other steps may be included in addition to the exemplified steps, other steps may be included except for some steps, or additional other steps may be included except for some steps.

Various embodiments of the present disclosure are intended to explain representative aspects of the present disclosure, rather than listing all possible combinations, and matters described in various embodiments may be applied independently or in combination of two or more.

In addition, various embodiments of the present disclosure may be implemented by hardware, firmware, software, or a combination thereof. For hardware implementation, one or more application specific integrated circuits (ASICs), digital signal processors (DSPs), digital signal processing devices (DSPDs), programmable logic devices (PLDs), field programmable gate arrays (FPGAs), It may be implemented by a processor (general processor), controller, microcontroller, microprocessor, or the like.

The scope of the present disclosure is software or machine-executable instructions (eg, operating systems, applications, firmware, programs, etc.) that cause operations according to methods of various embodiments to be executed on a device or computer, and such software or It includes a non-transitory computer-readable medium in which instructions and the like are stored and executable on a device or computer.

Claims (16)

In the memory management method, performed by a memory management device,
allocating at least one memory page to a local memory and a remote memory;
A process of checking whether a target performance ratio (α) required by the service is satisfied;
predicting an additional allocation size of the local memory in response to not satisfying the target performance ratio;
additionally allocating the local memory in consideration of the predicted additional allocation size;
The additional allocation size is,
calculating the number of remote memory accesses to be reduced in order to satisfy the target performance ratio, and setting the number of accesses to the remote memory to be reduced by the number of remote memory accesses to be reduced;
Whether or not the target performance ratio is satisfied,
It is determined based on a result of comparing the memory monitoring time period (P) and the value (L / α) divided by the application execution time (L) by the target performance ratio when the local memory is sufficient,
The additional allocation size is,
Divided memory management method, characterized in that calculated based on the function (T (n)) corresponding to the relationship between the number of accesses to the remote memory (n) and the application program execution time. delete According to claim 1,
The additional allocation size is,
A divided memory management method characterized in that it is predicted based on Equation 1 below.
[Equation 1]

In Equation 1, C is the additional allocation size, H(c) is the LRU distance histogram function, and Nn _t is the number of remote memory accesses to be reduced. delete According to claim 1,
The process of additionally allocating the local memory,
Checking the size of the remaining memory provided in the local memory;
Comparing the additional allocation size with the remaining memory size;
and additionally allocating the local memory in consideration of a comparison result between the additional allocation size and the remaining memory size. According to claim 1,
counting an access resumption time after access to the remote memory occurs; and
and reducing the size of the local memory by a predetermined size in response to the access resume time exceeding a predetermined time.
According to claim 1,
monitoring the utilization rate of the virtual processor;
and reducing the size of a memory page allocated to the local memory by a predetermined size in consideration of the utilization rate of the virtual processor.
In the memory management device,
a partitioned memory unit including a local memory included in a local area computing device and a remote memory included in a remote computing device;
a memory virtual machine unit that manages at least one memory page required for construction or execution of an application program;
a memory hypervisor unit provided between the divided memory unit and the memory virtual machine unit and allocating the at least one memory page to a local memory and a remote memory;
A performance predictor for determining whether a target performance ratio (α) required by the service is satisfied, and predicting an additional allocation size of the local memory in response to the target performance ratio not being satisfied;
The additional allocation size is,
calculating the number of remote memory accesses to be reduced in order to satisfy the target performance ratio, and setting the number of accesses to the remote memory to be reduced by the number of remote memory accesses to be reduced;
Whether or not the target performance ratio is satisfied,
It is determined based on a result of comparing the memory monitoring time period (P) and the value (L / α) divided by the application execution time (L) by the target performance ratio when the local memory is sufficient,
The additional allocation size is,
Divided memory management apparatus, characterized in that calculated based on the function (T (n)) corresponding to the relationship between the number of accesses to the remote memory (n) and the application program execution time. delete According to claim 8,
The performance prediction unit,
Partitioned memory management apparatus, characterized in that for estimating the additional allocation size based on the following equation (2).
[Equation 2]

In Equation 2, C is the additional allocation size, H(c) is the LRU distance histogram function, and Nn _t is the number of remote memory accesses to be reduced. delete According to claim 8,
The memory hypervisor unit,
The partitioned memory management apparatus of claim 1 , further allocating the local memory in consideration of the additional allocation size. According to claim 12,
The memory hypervisor unit,
Check the size of the remaining memory provided in the local memory,
Compare the additional allocation size with the remaining memory size;
The partitioned memory management apparatus of claim 1 , further allocating the local memory in consideration of a comparison result between the additional allocation size and the remaining memory size. According to claim 8,
The memory hypervisor unit,
and counting an access resume time after access to the remote memory occurs, and reducing the size of the local memory by a predetermined size in response to the access resume time exceeding a predetermined time.
According to claim 8,
The memory hypervisor unit,
The divided memory management apparatus, characterized in that for monitoring the use rate of the virtual processor, and reducing the size of a memory page allocated to the local memory by a predetermined size in consideration of the use rate of the virtual processor.
In the memory management device,
a partitioned memory unit including a local memory included in a local area computing device and a remote memory included in a remote computing device;
a memory virtual machine unit that manages at least one memory page required for construction or execution of an application program;
It is provided between the divided memory unit and the memory virtual machine unit, allocates the at least one memory page to a local memory and a remote memory, checks whether a target performance ratio (α) required by a service is satisfied, and In response to not satisfying the performance ratio, a memory hypervisor unit predicting an additional allocation size of the local memory and additionally allocating the local memory in consideration of the predicted size of the local memory;
The additional allocation size is,
calculating the number of remote memory accesses to be reduced in order to satisfy the target performance ratio, and setting the number of accesses to the remote memory to be reduced by the number of remote memory accesses to be reduced;
Whether or not the target performance ratio is satisfied,
It is determined based on a result of comparing the memory monitoring time period (P) and the value (L / α) divided by the application execution time (L) by the target performance ratio when the local memory is sufficient,
The additional allocation size is,
Divided memory management apparatus, characterized in that calculated based on the function (T (n)) corresponding to the relationship between the number of accesses to the remote memory (n) and the application program execution time.

Images