KR19980037118A - Predictive Technique-Based Buffer Replacement Using Neural Networks - Google Patents

Abstract

1. TECHNICAL FIELD OF THE INVENTION

A method of replacing a buffer in a multimedia service system.

2. The technical problem to be solved by the invention

Provides a buffer replacement method using neural networks for high performance service in multimedia database service system with data access biased over time.

3. Summary of Solution to Invention

First, it classifies the requested object into a hot set and a cold set. And in order to classify the required object, the past time interval values of this object pass through the neural network and divide the buffer into LFU area and LRU area to take advantage of LFU and LRU. The LFU area includes objects classified as a hot set and manages the buffer by the LFU method, and the LRU area includes objects classified as a cold set and manages the buffer by the LRU method.

4. Important uses of the invention

Applied to multimedia database systems supporting audio-on-demand (AOD) and video-on-demand (VOD) services.

Description

Predictive Technique-Based Buffer Replacement Using Neural Networks

Recently, multimedia services are increasing rapidly and will continue to increase. The database system that effectively manages this multimedia provides a buffer manager for quickly servicing the requested multimedia object. The present invention relates to a buffer replacement method of the above-described buffer manager for improving the performance of a multimedia database management system.

In other words, the present invention relates to a method of managing a buffer, which is one of resources most frequently accessed in a storage system, and a multimedia information service system in which a data access type changes with time. The present invention relates to a buffer replacement method for improving performance.

A buffer is a memory space used to hold frequently used data in main memory among data on a disk, which is a physical storage device. Among the modules constituting the storage system, a buffer management module is a module that manages a buffer in a memory and has the following functions. 1) Frequently used data is read from the disk into the main memory buffer which is immediately accessible. 2) Provides an operation method to place the data read from disk into the buffer if the newly requested data does not exist in the buffer. 3) The most ineffective data should be selected as the victim when replacing the buffer. This improves the overall performance of the storage system.

In many cases, the multimedia service system accesses data in a way that is different from that of a typical database. In real audio-on-demand (VOD) and video-on-demand (VOD) systems, some of the most popular objects are more likely to be retrieved in the future than others. As a result of analyzing the access type, objects are classified into two groups in the multimedia service system. One is that there are objects that will be queried more frequently in the future (hot set) and the other is that there will be objects that are less queried in the future (cold set). The access type of objects of considerable popularity is shown as an ascending phase in which the number of accesses increases rapidly over time and an extinction phase in which the number of accesses gradually decreases. In other words, over time, popular objects transition from a hot set to a cold set. This form of access is called time-varying skewed data access.

Various buffer replacement methods have been studied for database buffering, processor caching, and virtual memory paging. Methods such as first-in first-out (FIFO), last recently used (LRU), last frequently used (LFU), and modification of LRU (LRU-k) have been used in various fields. However, these buffer replacement methods are not suitable for the above-described approach, that is, buffer replacement in a service environment in which both biased and random data access forms exist over time.

Application of existing methods to the data access aspects of AOD and VOD analyzed in the present invention results in system degradation. Investigations of the present invention show that LRU and LRU-k are not suitable for a biased data approach. On the other hand, LFU is more efficient for biased data access, but the performance decreases as the degree of bias is alleviated. In addition, LFU is a property that depends on the frequency of access, which results in system degradation when an object transitions from a hot set to a cold set. Therefore, LRU or LFU is not suitable for multimedia service system where objects with both unbiased and random data access are present at the same time and the degree of bias and bias are variable over time.

Accordingly, an object of the present invention is to provide a prediction-based buffer replacement method using a neural network to predict data to be accessed in the future under the data access situation.

1 is a hardware configuration diagram to which the present invention is applied;

2 is a module configuration diagram of a storage system to which the present invention is applied;

3 is a buffer replacement structure to which the present invention is applied,

4 is a structural diagram of a buffer and a buffer element used in the buffer replacement method of the present invention;

5 is an operational flowchart of a buffer replacement method according to the present invention;

6 is a performance evaluation analysis diagram of the existing method.

* Explanation of symbols for the main parts of the drawings

21: user 22: interface management module

23: data storage / access management module 24: transaction management module

25: database

In order to achieve this object, the buffer replacement method according to the present invention uses a neural network technique to predict an access time interval of an object and a technique of mixing LRU and LFU to replace an actual buffer. The main role of the neural network is to classify the required objects into hot sets and cold sets. A hot set is a set of objects that have a great deal of popularity, with a biased data access, and a cold set is a set of objects that are not very popular and have random access. For this purpose, the past time interval values of the object are passed through the neural network. The present invention divides the buffer into an LFU region and an LRU region in order to take advantage of the LFU and LRU. The LFU area includes objects classified as a hot set and manages the buffer by the LFU method, and the LRU area includes objects classified as a cold set and manages the buffer by the LRU method.

Hereinafter, with reference to the accompanying drawings will be described an embodiment of the present invention;

1 is a configuration diagram of a hardware environment to which the present invention is applied.

1, Pi (1 = i = n) 11 is a processor, Mem 12 is a main memory memory area, SM 13 is a shared memory area in memory, and IOP 14 is dedicated to input and output. The processor and the DISK 15 are disks which are auxiliary storage devices. As such, the present invention applies to a general hardware environment including a single processor or multiple processors and a shared memory area.

2 shows a modular structure of a storage system in which the present invention is applied and operated. The storage system to which the present invention is applied basically stores an interface management module 22 connecting the user 21 and the system, and creates and processes a database 25 on a disk. It consists of an access management module 23, a transaction management module 24 that is responsible for concurrency control and recovery of transactions, and a buffer replacement method of the present invention. The buffer manager using the () is included in the data storage / access management module (23).

3 is a schematic structural diagram of buffer replacement in a storage system to which the present invention is applied.

The previous access information of the object requested by the user is input to the input layer 31 of the neural network, and the next access probability is calculated by the hidden layer 32 of the backpropagation neural network (BPN). predicted value to output layer, 33 Outputs When the prediction value is smaller than the threshold value, the required object is classified as a hot set, and when it is larger than this value, it is classified as a cold set. The object classified as a hot set is sent to the LFU region 35 of the buffer region 34 and managed by the LFU method, and the object classified as a cold set is sent to the LRU region 36 to the LRU. Managed by the method.

Therefore, the present invention obtains the access information of the object using neural network, and determines whether the object will continue to be accessed or is less likely to be accessed through this information, and based on this, LFU or LRU method useful for random access. The efficient buffer replacement mechanism is applied to each of them.

4 shows the structure of the buffer area 41 and object elements to be included in the buffer used in the present invention.

The buffer of a predetermined size is divided into an LFU area 42 and an LRU area 43, and the size of these areas is dynamically operated. That is, if there are many objects in the hot set, the LFU area will be relatively larger than the LRU area, and if the opposite occurs over time, the LRU area will be larger than the LFU area. The objects classified as hot sets and the objects classified as cold sets are linked by groups, and each group has a head pointer.

Information on object elements included in the buffer includes a buffer type (45) for distinguishing whether an object is a hot set or a cold set, and a past access including k arrival time prediction values. Information (Previous History Values) 46, Last Access Time 47 for applying the LRU Method, Access Frequence 48 for applying the LFU Method, Hot Set, Cold Set It consists of a Previous Buffer Pointer (49), a Next Buffer Pointer (50), and an Object Content (51) to form a ring by connecting buffer objects.

5 is an operation flowchart according to the buffer replacement method of the present invention, the operation is performed as follows.

1) When an object is requested from the user (501), the buffer manager first obtains previous access information of the requested object (502) and passes the access information to the BPNN, which is a neural network, to obtain a next access probability (prediction value) (503). The predicted value is compared with a threshold to classify the required object into a hot set or a cold set (504).

2) If the requested object exists in the buffer, the buffer type of the requested object is updated according to the result of 1), and the object is read from the buffer (505, 516).

3) As a result of 1), if the requested object is classified as a hot set, and this object does not exist in the buffer, it is checked whether there is an empty buffer (506, 507).

4) In 3), if there is an empty buffer, the object is placed in the empty buffer and the buffer is included in the LFU area (507, 511).

5) In 3), if there is no empty buffer, candidate victims are selected from buffers in the LFU region by the LFU method (508). The final victim is selected by applying the LRU method to the candidate victim and the buffers in the LRU region (509). After removing the victim from the buffer and putting the requested object into the buffer, the object is included in the buffer area (510).

6) As a result of 1), if the requested object is classified as a cold set and this object does not exist in the buffer, it is checked whether there is an empty buffer (512).

7) In 6), if an empty buffer exists, the object is put into the empty buffer and included in the LRU area (515).

8) In 6), if there is no empty buffer, the victim is selected by the LRU method (513,514).

As a result, the buffer replacement method proposed in the present invention divides the cache buffer into two areas, LRU and LFU, in order to use only the goodness of LRU and LFU, and each area is operated by the LRU and LFU methods. The neural network is used to classify the required object into hot set and cold set. Simulation results using the data access information of AOD in operation shows that the proposed method outperforms other buffer replacement methods such as LRU, LFU, and LRU-k.

The present invention showed better performance by experiment than the existing buffer replacement methods presented in other inventions or thesis.

In the first experiment, we simulate the random access to a set of objects with Zipf variance. This experiment considers the effect on buffer performance according to the degree of data access bias in using LRU and LFU. Experimental results are as shown in Figure 6, showing the average value of the results of 10 times. The simulation results show that LFU outperforms LRU as the data access bias increases. This is because LRU cache buffers can only be pushed out of the buffer because objects that are accessed once are always filled in the buffer line. On the other hand, LFU is suitable for biased data access because it has the property of distinguishing between frequently referenced and non-referenced objects. If the degree of bias is small, the difference in LRU and LFU hit rates is due to random data access patterns.

As a second experiment, the present invention used accessibility information of Korean songs that are being serviced by the Internet to model the biased data access of multimedia objects. The AOD system currently serves about 310 songs and has had 12,000 data accesses in the last month. With this data access information, the present invention evaluated the performance of LRU, LFU, LRU-k and our proposed method (OUR in Table 1). The results are shown in Table 1. According to Table 1, the experimental results show that the proposed method is superior to other methods. In particular, it can be seen that the proposed method is superior to other methods as the buffer size is smaller. This is because the proposed method improves the hit rate by accurately classifying objects into a hot set or a cold set through neural networks.

Buffer size LRU LRU-2 LRU-3 LFU OUR 20406080100 0.2500.3840.4730.5310.574 0.2550.3880.4760.5330.588 0.2590.4020.4790.5420.589 0.2090.3000.3790.4390.491 0.2980.4390.5270.5730.593

Claims (4)

In a buffer replacement mechanism in a multimedia service system having a data access type biased with time, A first step using neural networks to classify objects into hot and cold sets by predicting the access time intervals for which the requested object will be required next time: Dividing the buffer into LFU and LRU regions; A third step of managing an object classified as a hot set to an LFU area and managing the LFU method; And And a fourth step in which the object classified as a cold set is sent to an LRU area and managed by the LRU method. The method of claim 1, The first step is, If the object is requested from the user, the buffer manager obtaining previous access information of the requested object; Passing the access information to the neural network BPNN to obtain a next access probability (prediction value); And And comparing the predicted value with a threshold to classify the required object into a hot set or a cold set. The method of claim 1, The third step, Checking whether there is an empty buffer if an object classified as a hot set does not exist in the buffer; If there is an empty buffer, placing the object in an empty buffer, including the buffer in the LFU area, updating the buffer type, and reading the object from the buffer; If there is no empty buffer, selecting a candidate victim from among the buffers in the LFU region by the LFU method, and selecting the final victim by applying the LRU method to the buffers in the candidate victim and the LRU region; And Removing the victim from the buffer and putting the requested object into the buffer, including the object in the buffer area, updating the buffer type, and reading the object from the buffer. The method of claim 1, The fourth step, Checking whether there is an empty buffer if an object classified as a cold set does not exist in the buffer; If an empty buffer exists, inserting the object into the empty buffer, including the buffer in the LRU area, updating the buffer type, and reading the object from the buffer; Selecting a victim by the LRU method if there is no empty buffer; And Removing the victim from the buffer and putting the requested object into the buffer, including the object in the buffer area, updating the buffer type, and reading the object from the buffer.