KR20190100850A - Method for predicting energy consumption of heap memory object and memory system implementing the same - Google Patents

Abstract

A memory system for predicting amount of energy consumption of a heap memory object comprises: a memory device; and a memory profiler extracting access information when a heap memory object allocated to the memory device by an application program accesses the memory device if a profiling request is received from the application program and determining the amount of energy consumption predicted when the heap memory object is allocated to the memory device using the access information.

Description

Energy consumption estimation method of heap memory object and memory system implementing the same {METHOD FOR PREDICTING ENERGY CONSUMPTION OF HEAP MEMORY OBJECT AND MEMORY SYSTEM IMPLEMENTING THE SAME}

The present invention relates to a method for predicting energy consumption of a heap memory object and a memory system implementing the same.

The energy crisis in the data center is coming. Data center energy usage is increasing as large Internet server processes, such as Google and Facebook, and large research institutes run large, high-performance applications. According to a recent study, energy consumption in data centers in the United States accounted for about 3 percent of US annual electricity use in 2016, and data center energy consumption is steadily increasing, according to statistics from last year's consumption.

Data center power and energy consumption is between 20% and 48% in memory. On the other hand, methods used in the central processing unit to control the energy consumed in the memory has been applied, and in particular, methods for cutting off the power of the memory rank (memory rank) or to adjust the power consumption voltage and frequency of the memory.

However, these methods have limited energy that can be limited, such as performance degradation when the application is actively running, and limitations in reducing the number of ranks or voltages and frequencies that can cut power. In addition, these methods can only be applied at the system level, but can not be applied when there are different requirements at the application level.

Non-Volatile Memory technology may be effective in this memory energy consumption problem. Unlike volatile memory, nonvolatile memory, which has been actively developed recently, has a higher energy efficiency than volatile memory because it does not recharge the memory cell. In addition, non-volatile memory, unlike flash memory, supports byte addressability and has a number of advantages such as higher density than volatile memory.

However, the nonvolatile memory has a longer delay time than the volatile memory, and thus, the nonvolatile memory cannot be utilized as the main memory. To compensate for this, we have implemented a Hybrid Main Memory System that uses both volatile and nonvolatile memory for main memory, and allocates heap memory objects for applications to optimize performance and energy consumption. Research is being actively conducted. However, the problem common to heap memory object placement studies in the existing hybrid memory system is that the energy access pattern and non-volatile memory itself of different objects are not considered depending on the purpose and operation of the application. There is a problem that does not fully optimize.

In detail, the energy consumption of the heap memory object accessed in different patterns may vary according to the characteristics of the memory device to which the object is allocated. For example, objects that are accessed primarily by writing data to memory consume more energy in memory devices that consume more write power. Previous studies on memory object allocation in hybrid memory systems did not examine the various access patterns of these heap memory objects, and did not consider the characteristics of the nonvolatile memory constituting the hybrid memory system to accurately predict the amount of energy consumed by the object. There was a problem that could not.

An object of the present invention is to provide a technique for estimating the energy consumption of a heap memory object when a heap memory object allocated by an application program is allocated to the memory device.

When a memory system according to an embodiment of the present invention receives a profiling request from a memory device and an application program, a heap memory object allocated by the application program to the memory device accesses the memory device. And extracting access information and using the access information to determine an estimated energy consumption when the heap memory object is allocated to the memory device.

The access information includes at least one of a size of the heap memory object, a memory access size of the heap memory object, a memory usage of the heap memory object, or a lifetime of the heap memory object, and the memory profiler is configured to store the heap memory. Perform primary profiling on an object to extract at least one of the size of the heap memory object or the life time of the heap memory object, and perform secondary profiling on the heap memory object to store the memory of the heap memory object Extract at least one of the access size or the memory usage of the heap memory object.

The memory profiler determines energy information consumed according to an instruction executed on the memory device due to the heap memory object, and determines the energy consumption amount using the access information and the energy information.

When the memory device 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. do.

When the memory device is a non-volatile memory, the energy information may include at least one of energy consumption consumed by a preprocessing instruction, energy consumption consumed by a buffer access instruction, or energy consumption consumed by a write-back instruction. It includes one.

According to an embodiment of the present disclosure, a method of estimating energy consumption of a heap memory object by a memory profiler may include extracting access information for accessing a memory device by a heap memory object of the application when a profiling request is received from an application program. Determining energy information consumed according to an instruction executed on the memory device due to the heap memory object, and when the heap memory object is allocated to the memory device using the access information and the energy information. Determining an estimated energy consumption.

The access information includes at least one of a size of the heap memory object, a memory access size of the heap memory object, a memory usage of the heap memory object, or a life time of the heap memory object, and extracting the access information may include: Performing primary profiling on the heap memory object to extract at least one of a size of the heap memory object or a life time of the heap memory object, and performing secondary profiling on the heap memory object Extracting at least one of a memory access size of a heap memory object or a memory usage of the heap memory object.

When the memory device 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 device is a nonvolatile memory, the energy information includes at least one of energy consumption consumed by preprocessing instructions, energy consumption consumed by buffer access instructions, or energy consumption consumed by write-back instructions.

According to the present invention, in order to selectively allocate a heap memory object in consideration of the performance and power efficiency of the memory system, the energy consumption may be reduced while satisfying the overall performance of the memory system by estimating the energy consumption before actual allocation.

In addition, according to the present invention, in order to selectively allocate a heap memory variable in consideration of the difference in performance and power efficiency between the memory constituting the heterogeneous memory structure, while satisfying the overall performance of the heterogeneous memory by predicting the energy consumption before actual allocation It can reduce the energy consumed.

1 is a diagram illustrating a memory system according to an exemplary embodiment.
2 and 3 are diagrams illustrating an example memory device.
4 is a diagram illustrating a method of determining, by a memory profiler, access information of a heap memory object.
FIG. 5 is a diagram for describing a method of predicting an energy consumption of a heap memory object by a memory profiler according to an exemplary embodiment.

DETAILED DESCRIPTION Hereinafter, exemplary embodiments of the present invention will be described in detail with reference to the accompanying drawings so that those skilled in the art may easily implement the present invention. As those skilled in the art would realize, the described embodiments may be modified in various different ways, all without departing from the spirit or scope of the present invention. In the drawings, parts irrelevant to the description are omitted in order to clearly describe the present invention, and like reference numerals designate like parts throughout the specification.

Throughout the specification, when a part is said to "include" a certain component, it means that it can further include other components, without excluding other components unless specifically stated otherwise.

1 is a diagram illustrating a memory system according to an exemplary embodiment, FIGS. 2 and 3 are diagrams illustrating an exemplary memory device, and FIG. 4 is a diagram illustrating a method of determining, by a memory profiler, access information of a heap memory object. It is a figure explaining.

Referring to FIG. 1, the memory system 1000 includes a memory device 100 and a memory profiler 200.

The memory device 100 may be either volatile memory or non-volatile memory.

2 is a diagram illustrating a memory device according to one embodiment.

Referring to FIG. 2, the memory device 100 may be a volatile memory, and in this case, the memory device 100 may be a dynamic random access memory (DRAM).

In the case of DRAM, data is stored in a memory cell in the form of a voltage in a storage capacitor. When a read of a row of the memory is performed, voltage sharing between a bit line precharged with a predetermined voltage and a storage capacitor occurs.

In this case, data stored in the memory cell is destroyed, and this destructive read causes the DRAM to perform a recovery operation to rewrite the data sensed by the sense amplifier to the memory cell. have.

3 is a diagram illustrating a memory device according to another exemplary embodiment.

Referring to FIG. 3, the memory device 100 may be a nonvolatile memory. In this case, the memory device 100 may be a spin-transfer torque random access memory (STT-RAM).

STT-RAM has the highest performance and highest durability among nonvolatile memories. However, STT-RAM has a drawback in that the write power is large. Specifically, the energy consumed when writing to the memory array is more than twice as high as that of DRAM. However, this problem can be solved through a partial-write method that performs only a part of write-back of the row buffer for energy optimization of the STT-RAM, which will be described later.

Unlike DRAM, STT-RAM reads nondestructively, so there are separate row buffers and amplification detectors and operate independently of each other. Thus, when writing to the memory array is performed, the update occurs first in the row buffer. After that, when the address being accessed is not fetched into the row buffer, that is, when there is a conflict between the row buffers, the row buffer is rewritten to write to the actual memory array.

Meanwhile, the volatile memory and the nonvolatile memory may operate complementarily. Specifically, the volatile memory has a higher access speed than the nonvolatile memory, but has a disadvantage in that data is lost when the power is cut off. On the contrary, nonvolatile memory has lower access speed than volatile memory, but data is not lost even when power is cut off, and power efficiency is high. As a result, data center servers are increasingly using nonvolatile memory as an extension of memory or a storage cache of volatile memory, and have implemented and utilized heterogeneous memory systems that use them simultaneously. have.

When the memory profiler 200 receives a profiling request from the application 300, when the application 300 allocates a heap memory object to the memory device 100, the energy consumption of the allocated heap memory object is consumed. To predict.

Specifically, the type of the memory device 100 also affects the energy consumption, but the pattern in which the heap memory object used by the application 300 accesses the memory device 100 also affects the energy consumption. For example, in the case of a memory device in which the memory consumes a lot of energy to write data, a write-heap heap memory object may consume a lot of energy on the memory device.

Thus, in order to predict the energy consumption of the heap memory object, not only the characteristics of the memory device 100 itself but also the pattern in which the heap memory object approaches the memory device 100 must be considered.

Therefore, when the memory profiler 200 receives a profiling request from the application 300, the access information that the heap memory object allocated by the application 300 to the memory device 100 accesses the memory device 100. Extract

In this case, the access information includes at least one of the size of the heap memory object, the memory access size of the heap memory object, the memory usage of the heap memory object, or the life time of the heap memory object, and the memory profiler 200 performs profiling. By performing a plurality of times, the access information of the heap memory object is determined.

4 is a diagram illustrating a method of determining, by a memory profiler, access information of a heap memory object.

Referring to FIG. 4, the memory profiler 200 performs primary profiling on a heap memory object.

Primary profiling is the first pass, which is a fast run. In a quick implementation, the memory profiler 200 distinguishes different variables of the application 300 and extracts a hash value of a function call stack to map objects of each variable to the corresponding variable.

In detail, the memory profiler 200 intercepts a dynamic allocation function such as a malloc function by using a wrapper library, and then stores the call stack of the intercepted dynamic allocation function as a string and hashes it. The memory profiler 200 extracts the hash values of all the objects and writes them to the result file. After that, offline processing is performed to select main variables to perform secondary profiling and list the hash values in the file.

The memory profiler 200 performs primary profiling on the heap memory object to extract at least one of the size of the heap memory object or the life time of the heap memory object.

Thereafter, the memory profiler 200 performs secondary profiling on the heap memory object.

Secondary profiling is a second pass, which is a slow run. In a slow performance, the memory profiler 200 performs profiling of access information in an instruction unit by using a pin tool.

For example, the memory profiler 200 may profile the access to all heap memory objects at runtime using the Intel Pin tool, in which case the profiling results for the heap memory objects are first displayed. You can refer to the hash table created in the execution and save it as a result of the heap memory object to which the object belongs.

Meanwhile, in the present specification, the memory profiler 200 uses two-pass memory profiling, which is profiled through first and second executions, to extract access information of the heap memory object. Other memory profiling techniques may be used to extract the access information of.

The memory profiler 200 performs secondary profiling on the heap memory object to extract at least one of the memory access size of the heap memory object or the memory usage of the heap memory object.

The memory profiler 200 uses the access information to determine the estimated energy consumption when the heap memory object is allocated to the memory device 100.

In detail, the memory profiler 200 determines energy information consumed according to an instruction executed on the memory device 100 due to the heap memory object, and determines an energy consumption amount using the access information and the energy information.

In this case, the energy information may be determined differently according to the type of the memory device 100.

If the memory device 100 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.

For example, when the memory device 100 is DRAM among volatile memories, the ACT instruction that activates the banks and rows that are accessed before the memory read and write instructions are operated, and the data read or written is stored in the memory. PRE instructions, which precharge the bit lines to recover the array and prepare for subsequent ACT instructions, read and write instructions that perform actual memory read and write operations respectively, and periodically draw current to prevent data loss due to current leakage. A refresh command that supplies and recharges the memory device 100 may correspond to a command that consumes energy in the DRAM. In the present specification, the ACT instruction and the PRE instruction are defined as preprocessing instructions, and the energy consumed when accessing the memory array per byte by each instruction in the DRAM may be represented as shown in Table 1. Specifically, in Table 1, the energy consumed for the preprocessing and refresh instructions is the energy consumed when accessing the memory array per byte, and the energy consumed for the read and write instructions is per memory array and byte per byte. The energy consumed when accessing the row buffer.

command Energy consumed Preprocessing instruction 3.07 Read and write commands 1.19 Refresh instruction 0.35

On the other hand, the energy consumed described in Table 1 is merely an example, and may vary according to the standard and type of the memory device 100. When the memory device 100 is a volatile memory, the memory profiler 200 may include access information and By using the energy information, the energy consumption consumed when the heap memory object of the application program 300 approaches the memory device 100 may be estimated as shown in Equation 1 below.

In Equation 1, DE _i is the energy consumption estimated when the i-th heap memory object is allocated to the memory device 100, dE _{ACT + PRE} is the energy consumption per byte by the preprocessing instruction, dE _RW is read and write Energy consumption per byte by the instruction, dE _REF is energy consumption per byte by the refresh instruction, AV _i is the size of the memory access of the ith heap memory object, S _i is the size of the ith heap memory object, and T _i Is the life time of the i-th heap memory object.

That is, the memory profiler 200 may estimate energy consumption for each heap memory object, and add them all to estimate energy consumption of the entire heap memory object of the application 300.

If the memory device 100 is a nonvolatile memory, the energy information includes at least one of energy consumption consumed by preprocessing instructions, energy consumption consumed by buffer access instructions, or energy consumption consumed by write-back instructions. .

For example, if the memory device 100 is an STT-RAM of nonvolatile memory, like DRAM, preprocessing instructions for activating banks and rows and preparing for the next activation, buffer access instructions for accessing the row buffers, and between row buffers. A write-back instruction that writes back to a memory array in a row buffer when a collision occurs may correspond to an instruction that consumes energy in the STT-RAM. The energy consumed when accessing the memory array or row buffer per byte by each instruction in STT-RAM can be shown in Table 2. Specifically, in Table 2, the energy consumed for pre- and write-in instructions is the energy consumed when accessing the memory array per byte, and the energy consumed for the buffer access instruction accesses the row buffer per byte. When energy is consumed.

command Energy consumed Preprocessing instruction 2.68 Buffer access command 1.00 Write command 2.83

On the other hand, the energy consumed described in Table 2 is merely an example and may vary according to the standard and type of the memory device 100. When the memory device 100 is a nonvolatile memory, the memory profiler 200 may access information. And the energy information, the energy consumption consumed when the heap memory object approaches the memory device 100 by the application program can be predicted as shown in Equation 2 below.

In Equation 2, NE _i is the energy consumption estimated when the i-th heap memory object is allocated to the memory device 100, nE _{ACT + PRE} is the energy consumption per byte by the preprocessing instruction, nE _RBA is the buffer access instruction Energy consumption per byte by, nE _WB is energy consumption per byte by the write command, AV _i is the memory access size of the i th heap memory object, N _DC is the number of dirty cache blocks in the row buffer size , V _CB means the size of the cache block.

In order to count the number of cache blocks written to the memory array of the memory device 100, the memory profiler 200 may set a buffer having a size equal to the row buffer size of the memory device 100, and in the set buffer. N _DC may be determined by counting the number of corrupted cache blocks that occur. Specifically, when the average size of the heap memory objects to be profiled is larger than the row buffer of the memory device 100, the row buffer called when one workload is performed corresponds to a part of one object in most cases. Therefore, since the number of contaminated cache blocks in the virtual memory buffer equal to the row buffer size can be mapped to the number of contaminated cache blocks of the row buffer caused by one object, N _DC can be determined through this. In addition, the size of the cache block may be determined through hardware information of the memory device 100.

The memory profiler 200 may estimate energy consumption for each heap memory object, and add them all to estimate energy consumption of the entire heap memory object of the application 300.

FIG. 5 is a diagram for describing a method of predicting an energy consumption of a heap memory object by a memory profiler according to an exemplary embodiment.

Referring to FIG. 5, the memory profiler 200 receives a profiling request from the application program 300 (S100).

The memory profiler 200 extracts access information for the heap memory object of the application program 300 to access the memory device 100 (S110).

Specifically, the memory profiler 200 performs primary profiling on the heap memory object to extract at least one of the size of the heap memory object or the lifetime of the heap memory object, and the secondary profiling on the heap memory object. Extracts at least one of the memory access size of the heap memory object and the memory usage of the heap memory object. In this case, the access information includes at least one of the size of the heap memory object, the memory access size of the heap memory object, the memory usage of the heap memory object, or the life time of the heap memory object.

The memory profiler 200 determines energy information consumed according to an instruction executed on the memory device 100 due to the heap memory object (S120).

If the memory device 100 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.

If the memory device 100 is a nonvolatile memory, the energy information includes at least one of energy consumption consumed by preprocessing instructions, energy consumption consumed by buffer access instructions, or energy consumption consumed by write-back instructions. .

The memory profiler 200 determines the energy consumption predicted when the heap memory object is allocated to the memory device 100 using the access information and the energy information (S130).

According to the present invention, in order to selectively allocate a heap memory object in consideration of the difference in performance and power efficiency among memories constituting the heterogeneous memory structure, the energy consumption before the actual allocation is predicted to satisfy the overall performance of the heterogeneous memory. Energy can be reduced.

In addition, according to the present invention, in order to selectively allocate a heap memory variable in consideration of the difference in performance and power efficiency between the memory constituting the heterogeneous memory structure, while satisfying the overall performance of the heterogeneous memory by predicting the energy consumption before actual allocation It can reduce the energy consumed.

Although the embodiments of the present invention have been described in detail above, the scope of the present invention is not limited thereto, and various modifications and improvements of those skilled in the art using the basic concepts of the present invention defined in the following claims are also provided. It belongs to the scope of rights.

Claims (9)

As a memory system,
A memory device, and
Upon receiving a profiling request from an application, a heap memory object allocated by the application to the memory device extracts access information for accessing the memory device, and uses the access information to extract the access information. Memory profiler, which determines the estimated energy consumption when an object is assigned to the memory device
Memory system comprising a. In claim 1,
The access information includes at least one of a size of the heap memory object, a memory access size of the heap memory object, a memory usage of the heap memory object, or a lifetime of the heap memory object,
The memory profiler
Perform primary profiling on the heap memory object to extract at least one of a size of the heap memory object or a lifetime of the heap memory object,
And performing at least one of a memory access size of the heap memory object or a memory usage of the heap memory object by performing secondary profiling on the heap memory object. In claim 1,
The memory profiler
Determining energy information consumed according to an instruction executed on the memory device due to the heap memory object, and determining the energy consumption amount using the access information and the energy information. In claim 3,
If the memory device is a volatile memory (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. In claim 3,
If the memory device is a non-volatile memory (Non-Volatile Memory),
The energy information includes at least one of energy consumption consumed by preprocessing instructions, energy consumption consumed by buffer access instructions, or energy consumption consumed by postwrite instructions. The memory profiler predicts the energy consumption of heap memory objects.
Receiving a profiling request from an application, extracting access information for the heap memory object of the application to access a memory device,
Determining energy information consumed in accordance with instructions executed on the memory device due to the heap memory object, and
Determining an energy consumption amount estimated when the heap memory object is allocated to the memory device using the access information and the energy information.
Energy consumption prediction method comprising a. In claim 6,
The access information includes at least one of a size of the heap memory object, a memory access size of the heap memory object, a memory usage of the heap memory object, or a lifetime of the heap memory object,
Extracting the access information
Performing primary profiling on the heap memory object to extract at least one of a size of the heap memory object or a life time of the heap memory object, and
Extracting at least one of a memory access size of the heap memory object or a memory usage of the heap memory object by performing secondary profiling on the heap memory object. In claim 6,
If the memory device is volatile memory,
And 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. In claim 6,
If the memory device is a nonvolatile memory,
And the energy information comprises at least one of energy consumption consumed by a preprocessing instruction, energy consumption consumed by a buffer access instruction, or energy consumption consumed by a write-back instruction.

Images