KR100794969B1 - A method for adaptively retrieving data for load shading in clustered video servers - Google Patents

Abstract

A method for searching data adaptively to share a load in clustered video servers is provided to share the load among clusters without migrating or replicating the data by using adaptive data search changing length of rounds dynamically. A value of a disk bandwidth usage and a buffer usage is calculated from all clusters, and a set of the disk bandwidth usage and the buffer usage is stored. The set of the disk bandwidth usages is found when a selection parameter is the number of divisors of an available round length set. The disk bandwidth usage and the buffer usage for the predetermined cluster are calculated again when a customer requests a video stream stored in the cluster. It is checked whether the round length can be extended for the cluster. It is checked whether the round length can be extended without jitter. It is checked whether the buffer usage of which the round length is extended is smaller than one when the round length can be extended. The round length is extended when a buffer space is enough.

Description

Adaptive Data Retrieval for Load Sharing in Clustered Video Servers {A METHOD FOR ADAPTIVELY RETRIEVING DATA FOR LOAD SHADING IN CLUSTERED VIDEO SERVERS}

1 is a diagram showing an example of a configuration of a video server in which the number of clusters is two, and each cluster includes four homogeneous disks.

이며, 시간 (m + 2) * fr_j에서 라운드 길이가 fr_j에서 fr_{j +1}로 변화하는 경우의 스트림들의 검색 및 재생을 도시하는 도면.2 is Q = 1,

, And the time (m + 2) * a diagram showing a search and playback of a stream in the case of changes in the length of the round fr fr _j from _j to _{j +1} fr.

이며, 시간 (m + 2) * fr_j에서 라운드 길이가 fr_{j +1}에서 fr_j로 감소하는 경우의 데이터 검색 및 재생을 예시하는 도면.3 is Q = 1,

, And the time (m + 2) * a view illustrating a round length of data retrieval and reproduction in the case of reduced fr fr _j from _{j +1} in fr _j.

4 illustrates an algorithm specifically implementing an adaptive data retrieval method according to the present invention.

FIG. 5 shows how the proportion of customers under the GCP (Gradual Change of Popularity) scenario, i.e., the authorization ratio, depends on the arrival rate of the customer requests.

6 and 7 show how the authorization ratio depends on the value of α under the DCP (Rapid Change of Popularity) scenario when the arrival rate is 20 and 30 requests / minute.

<Explanation of symbols for the main parts of the drawings>

: 클러스터 k의 i번째 디스크

: I-th disk in cluster k

S _{i , m} : m-th segment of video V _i

: 서브-세그먼트

Sub-segment

Data read: read data

S _m : segment m

Playback of S _m : Playback for segment S _m

CV: traditional data retrieval method that does not allow adaptive round length adjustment

ADR: Adaptive Data Retrieval Method According to the Present Invention

The present invention is an adaptive data retrieval method for load sharing in a clustered video server, and more specifically, adaptive data capable of achieving load sharing between clusters without moving or copying data by dynamically changing round length. It is about a search method.

Recent advances in multimedia and network technologies have made it possible to provide customers with video-on-demand "VOD" services. To be able to service VOD, a video server must have the ability to deliver video to thousands of customers. Since video data requires high bandwidth and large storage space, video servers are typically built on disk arrays that can consist of hundreds of disks. A major concern in the design of the server is how to locate and retrieve video data to maximize the number of concurrent streams.

As a data retrieval technique for supporting a large number of customers, a round-based scheduling technique is generally used. In round-based scheduling, time is divided into equally sized periods called rounds, with each authorized customer being served once in each round. To ensure continuous playback, the data that needs to be played back is read from the disc in the previous round, while the data to be played back in the next round is read out in the current round. Otherwise, playback may be distorted or pauses may occur due to a timeout violation of the video data, which is referred to as jitter.

To effectively use disk bandwidth, video objects are often divided into segments and distributed over multiple disks. This scheme is called striping, where a segment is the maximum amount of consecutive data stored on one disk. Striping across too many disks involves additional complexity, which can lead to reliability problems. Therefore, it is realistic to divide one disk array into several clusters to limit the range of striping, to have each cluster independently form a striping group, and then to have each video stripe in one cluster.

In general, customers' requests tend to be excessively skewed to a few popular videos, so clusters containing copies of those popular videos are more likely to receive overload than other clusters. Since it is difficult to accurately predict the probability of access to the video that will be requested by future customers, it is impossible to place the video data in a balanced manner. Even more difficult is that the request patterns may change over time. For example, children's videos are often popular in early evenings, but are rarely requested at late nights.

To achieve load balancing in a clustered video server, existing approaches use dynamic data migration or replication techniques. The main idea of these schemes is to move or copy video data between clusters to achieve some degree of load balancing. However, these schemes require more resources, such as additional disk bandwidth and storage space, to move or copy, which can cause downtime or affect the performance of applications accessing the VOD system. In addition, this shift needs to be based on an assessment of the future demands of customers, which inaccurate predictions may significantly degrade system performance. Therefore, these approaches are not very suitable for video servers.

The present invention is derived from the above recognition, and aims to propose an adaptive data retrieval method capable of achieving load sharing between clusters without moving or copying data.

According to an aspect of the present invention for achieving the above object, an adaptive data retrieval method for load sharing in a clustered video server is as follows.

The transfer rate is tr, the disk seek time is T _s , each constant, the buffer size is B,

The time is divided into equally sized periods called rounds, and round-based scheduling is used where each authorized customer is served once in each round,

The video stream V _i is divided into a finite number of consecutive segments, each segment being subdivided into NS sub-segments (where each sub-segment has a length BR of data retrieved during the basic round). Size), each segment is stored on disks in a round-robin fashion, and the data playback speed is dr _i , the unit of which is bits / sec,

이 상기 비디오 스트림 V_i를 요청하고 있는 클러스터링된 비디오 서버에서,Consisting of C clusters, each cluster consisting of Q homogeneous disks, dividing the customers into C customer groups (CG ₁ , ..., CG _C ), and then the customers in group CG _k from the corresponding cluster k Configured to receive streams, and

In the clustered video server requesting this video stream V _i ,

When the disk bandwidth utilization DS _k (j) and the buffer utilization BS _k (j) for the cluster k when the round length is fr _j are given by Equations 1 and 2, respectively,

)(here,

)

A method of adaptively retrieving data for load sharing in a clustered video server,

(1) calculating the values of DS _k (j) and BS _k (j) for all clusters k (k = 1, ..., C) and maintaining their respective sets;

(2) Disk utilization DS _m when the selection parameter SL _m is ND indicating that the SL _m th element of the set FS containing the possible round lengths in ascending order, fr _SLm is selected as the round length for cluster _m Obtaining a set SM of (ND) (m = 1, ..., C);

(3) If a customer requests a video stream stored in cluster p, assume that a new customer will be authorized j = 1,... Recalculating the utilizations DS _p (j) and BS _p (j) for the cluster p for ND, where ND is the number of divisors for the NS;

(4) determining whether the selection parameter SL _p for the cluster p is SL _p ≤ ND-1 to determine whether the round length can be extended by 1 for the cluster p;

(5) Checking whether the disk bandwidth utilization DS _p (SL _p + 1) satisfies the following Equation 3 when increasing the selection parameter for the cluster p to determine whether it is possible to extend the round length without jitter. Doing;

에 대한 최대 서비스 시간을 나타냄)Where ε is the customer belonging to customer group CG _p

Maximum service time for)

(6) If it is determined in step (4) to (5) that the round length is expandable without jitter, whether the buffer utilization in the case of extending the round length for the cluster p is smaller than 1 Checking by to determine whether there is sufficient buffer space;

(7) if it is determined in step (6) that the buffer space is sufficient, increasing SL _p to SL _p + 1 to expand the round length;

(8) selecting element DS _l (ND) having the smallest value in SM and removing it from SM;

(9) checking whether the disk bandwidth utilization DS _l (ND-1), which is a case where the round length is reduced for the cluster l, is less than or equal to 1 to determine whether the disk bandwidth utilization is satisfied; And

(10) If it is determined in step (9) that DS _l (ND-1) ≤ 1, SL _{l is set} to ND-1 to reduce the round length, otherwise step (8) to (8) while SM is not empty Repeat step (9)

It characterized by including the.

In addition, a method for adaptively retrieving data for load sharing in a clustered video server according to another aspect of the present invention,

(11) if the customer closes the video stream stored in cluster k, recalculating DS _k (j) and BS _k (j) (j = 1, ..., ND);

(12) determining whether the following Equation 5 is satisfied;

(SL_k = 2, …, ND)

(SL _k = 2,…, ND)

(13) if it is determined in Equation 12 that Equation 5 is satisfied, determining whether Equation 6 is satisfied to determine whether reducing the round length would violate the disk bandwidth constraint; And

(14) if it is determined in step (13) that the disk bandwidth constraint is not violated, reducing SL _k to SL _k -1 to reduce the round length of cluster k;

Characterized in that it further comprises.

Hereinafter, with reference to the accompanying drawings, it will be described in detail an embodiment of the present invention.

1. System Model

라고 표시하기로 한다.The video server used in the present invention divides one disk array into clusters, and each cluster independently forms a striping group. The video server is divided into C clusters, and each cluster may be composed of Q homogeneous disks. To refer to the i th disk of cluster k

Will be displayed.

1 shows an example configuration of a video server in which the number of clusters is two, and each cluster includes four homogeneous disks. The video V _i is divided into a finite number of consecutive segments _{Si, 1} , _{Si, 2} ,... In FIG. 1, V ₁ is stored in cluster 1 and V ₂ is stored in cluster 2.

When a customer requests a video stream, the server allocates a certain amount of buffer space to store a portion of the video. The server also allocates disk bandwidth for each stream to continuously retrieve video data from the disks. To reduce the inevitable overhead of disk seek time, SCAN scheduling is used, where the disk head scans back and forth along the surface of the disk and searches for blocks it passes by.

2. According to the round length resource  Change in utilization

The choice of round length is important because it substantially determines the maximum number of concurrent users. Since retrieving large amounts of data in one round reduces the impact of disk latency, long rounds improve the efficiency of disk utilization. On the other hand, long rounds increase the buffer requirements because a lot of buffer space is required to store the retrieved data.

In the adaptive data retrieval method proposed in the present invention, each cluster has a unique round length and a buffer space. In order to overcome the stress on the disk bandwidth caused by the overloaded cluster, the method according to the invention increases the round length. This increases the requirement for buffer space, but can take up buffer space from less used clusters. However, the total buffer space is a limited resource, and furthermore, the process of changing the round length may introduce jitter. In order to effectively perform this balancing act, it is first necessary to know how resource utilization varies with round length.

(n = 1, …, NS)로 분할하는데, 여기서 각각의 서브-세그먼트의 크기는 길이 BR의 기본 라운드 동안 검색되는 데이터 크기에 해당한다. NS개의 서브-세그먼트들은 인접하게 저장되어 하나의 세그먼트를 구성하며, 세그먼트들은 도 1에 도시된 바와 같이 라운드-로빈(round-robin) 방식으로 배치된다(여기서 라운드-로빈 방식이란 그룹 내에 있는 모든 요소들을 한 번씩 순서대로 돌아가면서 처리하는 방법을 의미함).Since the data retrieved during a new round may not be stored contiguously, changing the round length may incur additional search overhead. To solve this, each data segment _{Si, m} is NS sub-segment

(n = 1, ..., NS), where the size of each sub-segment corresponds to the data size retrieved during the basic round of length BR. The NS sub-segments are contiguously stored to form one segment, and the segments are arranged in a round-robin manner as shown in FIG. 1 (where all elements within the group are round-robin schemes). How to process them in turn once in a while).

을 액세스한다. 만약 NS = 6이고,

부터

까지 6개의 서브-세그먼트가 하나의 세그먼트 S_1,3를 구성한다고 하면, 서버는 한 라운드 동안 4개의 인접한 서브-세그먼트

,

,

,

를

로부터 검색할 것이다. 그러나

와

는

상에 저장되어 있고,

와

는

상에 저장되어 있기 때문에, 다음 라운드 동안, 서버는 2개의 디스크

와

에 액세스할 필요가 있을 것이다.We will use the notation dv _j (j = 1,…, ND) to denote the jth divisor of NS, where ND is the number of divisors of NS. FS, the set of possible round lengths, is fr _j | fr _j = BR * dv _j . Assume that the elements of FS are sorted in ascending order. For example, if NS = 6, FS is fr ₁ = BR, fr ₂ = 2BR, fr ₃ = 3BR, fr ₄ = 6BR. Round lengths that do not correspond to FS are not allowed because they may require two searches to perform one read. For example, suppose 4BR is selected as the round length, and the server uses the layout in FIG. 1 to retrieve S _{1 , 3} .

To access it. If NS = 6,

from

Up to six sub-segments up to one segment S _1,3 , the server has four contiguous sub-segments in one round.

,

,

,

To

Will search from. But

Wow

Is

Stored in the

Wow

Is

Because it is stored on the server, during the next round, the server

Wow

You will need to access.

의 판독 시간을 발생시키는데, 여기서 tr은 디스크의 데이터 전송 속도이다. 결과적으로, 비디오 스트림 V_i를 서비스하는 것은

의 서비스 타임을 증가시킨다. 고객

이 비디오 스트림 V_i를 요청한다고 가정하자. 고객들은 C개의 클라이언트 그룹(CG₁, …, CG_C)으로 분할될 수 있는데, 여기서 그룹 CG_k 내의 고객들은 클러스터 k로부터 스트림을 수신한다. 만약 라운드 길이가 fr_j이면, 클러스터 k에 대한 전체 서비스 타임 ST_k(j)를 다음 수학식 7과 같이 얻을 수 있다.Let's see how the buffer and disk bandwidth requirements for video V _i with a data rate of dr _i (bits / sec) are affected by the round length fr _j (j = 1,…, ND). A typical seek time model is used, which requires a constant search overhead of T _s (search time + rotational delay) for one read of adjacent data. Searching for the video stream V _i is equivalent to the overhead of T _s

Generates a read time, where tr is the data transfer rate of the disk. As a result, servicing video stream V _i

To increase the service time. customer

Suppose we request this video stream V _i . Customers can be divided into _C client groups CG ₁ ,..., CG _C , where customers in group CG _k receive a stream from cluster k. If the round length is fr _j , the total service time ST _k (j) for the cluster k can be obtained as in Equation 7 below.

For ease of explanation, assume that disk loads are distributed evenly among the disks in the same cluster. Such load balancing can be easily accomplished by delaying authorization for customers. Bandwidth utilization for a disk is often defined as the ratio of overall service time to round length. Since there are Q disks in one cluster, the disk bandwidth utilization for cluster k can now also be determined as shown in Equation 1 below with reference to DS _k (j).

<Equation 1>

만큼 증가시킨다. 이제 클러스터 k에 대한 버퍼 이용도 BS_k(j)를 앞서 언급한 다음 수학식 2와 같이 얻을 수 있다.Let B be the total buffer size. Since double buffering is used for SCAN scheduling, servicing video stream V _i reduces buffer utilization.

Increase by. The buffer utilization for cluster k can now be obtained as shown in Equation 2 below, BS _k (j).

<Equation 2>

3. Jitter  Condition for no round adjustment

Increasing the round length may prevent the data block from reaching on time, which may result in jitter in some video streams. This happens because too few data blocks could be read to support the new round length.

이며, 시간 (m + 2) * fr_j에서 라운드 길이가 fr_j에서 fr_{j +1}로 변화하는 경우의 스트림들의 검색 및 재생을 도시한다. 지터가 없는 서비스를 유지하기 위해, 한 라운드 동안 검색된 데이터는 다음 라운드 동안 재생을 위해 사용될 수 있어야 한다. 예를 들어, 도 2(a)에서, 시간 m * fr_j 및 시간 (m + 1) * fr_j 사이에 소모된 데이터의 모든 아이템은 시간 m * fr_j에서 버퍼 내에서 사용 가능하게 될 필요가 있다. 도 2(a)는 라운드의 길이를 늘이는 것이 현재 진행 중인 스트림들에서 어떻게 지터를 발생시킬 수 있는지를 예시한다. 도면으로부터, 스트림들 S3 및 S4에 의해 요청되는 블록들이 시간 (m + 3) * fr_j에서 사용할 수 없기 때문에 현재 진행 중인 스트림들에 대하여 시간 (m + 3) * fr_j부터 시간 (m + 4) * fr_j까지 데이터가 없는 것을 관찰할 수 있다.2 is Q = 1,

, And it illustrates the retrieval and playback of a stream in the case of a round length change in fr fr _{j +1} in _j in the time (m + 2) fr * _j. In order to maintain a jitter free service, the data retrieved during one round must be available for playback during the next round. For example, in FIG. 2 (a), all items of data consumed between time m * fr _j and time (m + 1) * fr _j need to be made available in the buffer at time m * fr _j . have. 2 (a) illustrates how increasing the length of the round can cause jitter in currently ongoing streams. Since from the figure, streams S3 and the block requested by the S4 to time (m + 3) * not available on the fr _j with respect to the stream that is currently proceeding time (m + 3) * fr _j from the time (m + 4 ) * You can observe that there is no data up to fr _j .

2 (b) illustrates a situation where expansion of the round length does not cause jitter. This is because all items of data that need to be played back between time (m + 3) * fr _j and time (m + 5) * fr _j are available at time (m + 3) * fr _j . From these observations, it is easy to see that in order to ensure jitter-free operation, the service time for the new round length fr _{j + 1} in a single disk cluster should not exceed fr _j . The Q of the disk exists, since the disc loading is uniformly distributed among the disks in the same cluster, the conditions for the round extension without a jitter to fr _{j + 1} from fr _j is obtained as shown in Equation 8 (j = 1,…, ND-1).

Using Equation 1, the inequality can be rewritten as Equation 9 below.

The server may need to reduce the round length to get buffer space. Since buffer space is obtained at the expense of disk bandwidth, the server needs to check disk bandwidth constraints. That is, reducing the round length from fr _{j + 1} to fr _j should satisfy DS _k (j) ≦ 1 (j = 1,…, ND−1).

이며, 시간 (m + 2) * fr_j에서 라운드 길이가 fr_{j +1}에서 fr_j로 감소하는 경우의 데이터 검색 및 재생을 예시한다. 도 3으로부터, 다음과 같은 점들을 관찰할 수 있다. (1) 감소된 라운드 길이 fr_j에 대한 데 이터 검색이 시간 (m + 3) * fr_j에서 시작한다는 점과 (2) 라운드 길이를 fr_j+1로부터 fr_j로 감소시키는 것은, 시간 (m + 4) * fr_j과 시간 (m + 5) * fr_j 사이에 재생되는 모든 데이터가 시간 (m + 4) * fr_j에서 사용 가능하기 때문에, 지터를 발생시키지 않는다는 점을 관찰할 수 있다.3 is Q = 1,

, And it illustrates the data search and reproduction is performed when the round length reduced to fr fr _{j +1 j} at the time _{(m + 2) * fr j} . From FIG. 3, the following points can be observed. (1) the data search for reduced round length fr _j starts at time (m + 3) * fr _j and (2) reducing the round length from fr _{j + 1} to fr _j is time (m It can be observed that since all data reproduced between + fr _j and time (m + 5) * fr _j are available at time (m + 4) * fr _j , it does not cause jitter.

4. Adaptive Data Retrieval Scheme

Let SL _k be a selection parameter indicating that the SL _k th element of the FS, fr _SLk is selected as the round length for the cluster k. For example, if FS = BR, 2BR, 3BR, 6BR and SL _k = 2, 2BR is selected as the round length. From Equations 1 and 2, it can be readily seen that higher values of fr _j decrease DS _k (j) but increase BS _k (j). From this observation, the ADR scheme adjusts the round length in the following manner.

(1) In order to meet the requirements of clusters overloaded with disk bandwidth, a relatively high value of SL _k is assigned to these clusters,

(2) To obtain buffer space to increase the round length of overloaded clusters, a relatively low value of SL _k is allocated to lightly loaded clusters.

The difficulty here is to accurately determine when it is possible to change the round length in a jitter free manner. The present invention uses a threshold-based approach in which a change in round length is caused by a change in disk bandwidth utilization. In practice, this means that the round length can be changed if the customer requests or closes the video stream. Explain what happens in each case.

4.1 Customer Video Stream  If requested

The server maintains a set of utilizations DS _k (j) and BS _k (j) for all clusters k (k = 1,…, C). When a customer requests a video stream stored in cluster p, the server assumes that a new customer will be authorized and j = 1,... Recalculate the utilizations DS _p (j) and BS _p (j) for cluster p for ND. The server then extends the round length if necessary, then checks whether there is enough disk bandwidth and buffer space for the new customer.

인지를 체크한다(여기서, ε는 고객 그룹 CGp에 속하는 고객

에 대한 최대 서비스 시간, 즉

을 나타냄)(제5행). 수학식 9로부터, 상기 조건이 지터가 없는 라운드 확장을 보장한다는 것을 알 수 있다. 상기 조건이 만족되는 경우, 서버는 충분한 버퍼 공간이 라운드 확장을 위해 사용 가능한지를 체크한다(제6행). 버퍼 공간이 충분한 경우, 서버는 SL_p의 값을 1만큼 증가시킨다(제7행). 그렇지 않으면, SL_p의 값은 변화하지 않는다.4 is a diagram illustrating an algorithm that specifically implements an adaptive data retrieval method according to the present invention. For ease of explanation, it is assumed that ND ≧ 2. First, the server must confirm that SL _p is smaller than ND, ie the round length can be extended. If the disk bandwidth utilization for a particular cluster exceeds the threshold, the server extends the round length to reflect the increased load. For this purpose, the server

Check whether the ε is a customer belonging to the customer group CGp

Maximum service time for, i.e.

(Line 5). From Equation 9, it can be seen that the condition ensures jitter-free round expansion. If the condition is met, the server checks whether enough buffer space is available for round expansion (line 6). If there is enough buffer space, the server increments the value of SL _p by one (line 7). Otherwise, the value of SL _p does not change.

및

와 같이 이미 만족하기 때문에 불필요하다. 인가에 대해 결정하기 위하여, 서버는 먼저

인지 여부를 검사한다(제10행). 제10행의 조건을 위반할 경우, 새로운 고객은 디스크 대역폭의 부족 때문에 거절된다. 아니라면, 서버는

인지 여부를 검사한다(제11행). 이 조건을 만족하면, 새로운 고객이 인가된다. 그렇지 않으면, 서버는 한 클러스터의 라운드 길이를 감소시킴으로써 새로운 고객에 대한 버퍼 공간을 획득할 수 있다. 가장 큰 라운드들(즉, SL_k = ND)이 가장 많은 버퍼 공간을 사용하기 때문에, 본 발명에 따른 방법은 이들을 가지고 클러스터들을 테스트한다. SM을 SL_m이 ND인 디스크 이용도 DS_m(ND)(m = 1, …, C)의 집합이라고 하자. SM이 공집합이라면, 새로운 고 객은 버퍼 공간의 부족 때문에 거절된다. 아니라면, SM 내에서 가장 가볍게 부하가 걸린 클러스터를 찾기 위하여, 서버는 SM 내에서 DS_l(ND)의 가장 작은 값을 찾아서, SM으로부터 제거한다(제15-16행). 이것이 디스크 대역폭 제약을 위반하지 않는다면(제17행), 서버는SL_l의 값을 ND로부터 ND-1로 감소시킨다(제18-19행). 라운드 길이에서의 이러한 감소로 인해 새로운 고객이 수용 가능하게 된다.Next, the server checks whether a new customer can be authorized, which means that if the round length is extended,

And

This is unnecessary because it is already satisfied. To determine authorization, the server first

Check whether it is (line 10). If you violate the conditions in line 10, new customers will be rejected due to lack of disk bandwidth. If not, the server

Check whether it is (line 11). If this condition is met, a new customer is authorized. Otherwise, the server can free up buffer space for new customers by reducing the round length of one cluster. Since the largest rounds (ie SL _k = ND) use the most buffer space, the method according to the invention tests the clusters with them. Assume SM is a set of DS _m (ND) (m = 1, ..., C) for which disk _m is ND. If the SM is empty, new customers are rejected due to lack of buffer space. If not, to find the lightest loaded cluster in the SM, the server finds the smallest value of DS _l (ND) in the SM and removes it from the SM (lines 15-16). If this does not violate the disk bandwidth constraint (line 17), the server decreases the value of SL _l from ND to ND-1 (lines 18-19). This reduction in round length makes new customers acceptable.

4.2 Customer Video Stream  If closing

(SL_k = 2, …, ND) 여부를 체크한다. 이 조건이 만족되면, 서버는 라운드 길이를 감소시키는 것이 디스크 대역폭 제약을 위반시킬 것인지(즉,

) 여부를 체크한다. 이 조건이 유지될 수 있다면, 서버는 SL_k의 값을 SL_k-1로 감소시킨다.By reducing the round length in a timely manner, the server can gain buffer space, which can then be used to extend the round length of overloaded clusters. To this end, when the customer closes the video stream stored in cluster k, the server recalculates DS _k (j) and BS _k (j) (j = 1,…, ND),

(SL _k = 2, ..., ND) is checked. If this condition is met, the server can determine if reducing the round length would violate disk bandwidth constraints (i.e.

Check it. If this condition can be maintained, and the server reduces the value of SL _k in SL _k -1.

5. Experimental Results

라고 하자. 다음과 같은 2가지 서비스 시나리오를 고려한다.To evaluate the effectiveness of the present invention, a server with 40 IBM Ultrastar36Z15 disks with tr and T _s of 55 MB / s and 11.2 ms, respectively, was simulated. The server is divided into ten clusters, with each cluster consisting of four disks. The arrival of the customer request is assumed to follow the Poisson distribution. In addition, it is assumed that all video is 90 minutes long and has the same bit-rate of 1.5 Mbps typical for MPEG 1 playback. NS is set to 6, and FS is BR = 0.5, 2BR = 1, 3BR = 1.5, and 6BR = 3. The total buffer size B is 2 GB. The cluster position of each movie is chosen arbitrarily. p _i (t) is the access probability of video V _i (i = 1,…, NV) at time t, where NV is the number of movies,

Let's say Consider two service scenarios:

(1) gradual change of popularities (GCP)

). θ = 0.0으로 설정하는데, 이것은 실제 VOD 애플리케이션에 대한 실 측정값에 해당한다. 처음에, p₁(0) > … > p_NV(0)이며, 시간 τ에서,

이다. 영화의 인기도는 τ의 주기로 시간에 대해 순환하는데, 예컨대

이다. 이러한 패턴은 하루의 서로 다른 시간들에서 관객의 유형들을 변화시키는 것에 해당할 수 있다.Access probability follows the Zipf distribution

). Set θ = 0.0, which corresponds to the actual measurement for the actual VOD application. Initially, p ₁ (0)>... > p _NV (0) and at time τ,

to be. The popularity of a movie circulates over time in a period of τ, for example

to be. This pattern may correspond to changing the types of audience at different times of the day.

(2) dramatic change of popularities (DCP)

In this case, the popularity of a particular movie increases rapidly up to α for every period of τ. Set p ₁ (0) = α. The popularity of the movies then circulates in a cycle of τ. For example, at time τ, p ₂ (τ) increases to α. This scenario corresponds to the periodic distribution of a new movie or drama.

We will compare the method according to the invention with three existing methods, CV1, CV2 and CV3, which do not allow for adaptive round length adjustment. The round lengths of CV1, CV2 and CV3 are assumed to be 1, 1.5 and 3 seconds, respectively. After that, we examined the number of authorized customers for 20 hours under the GCP and DCP scenarios. The value of τ is assumed to be 30 minutes.

5 shows how the proportion of customers under the GCP (Gradual Change of Popularity) scenario, i.e., the authorization ratio, depends on the arrival rate of the customer requests. In this case, the method ADR according to the invention showed better performance than existing schemes under all workloads, authorizing more customers between 1% and 28%.

6 and 7 show how the authorization ratio depends on the value of α under the DCP (Rapid Change of Popularity) scenario when the arrival rate is 20 and 30 requests / minute. 6 and 7, the method ADR according to the invention shows the best performance under all workloads. The method according to the invention authorizes more customers between 21% and 33% than CV1, authorizes more customers between 10% and 21% than CV2, and more customers between 1% and 19% than CV3. Was applied. From the figures it can be observed that the performance difference between the method according to the invention and CV2 increases as the value of α increases, which can be explained as follows. In other words, as the value of α increases, the requests are concentrated in popular clusters so that the disk bandwidth is saturated in a shorter time. It is also observed that CV3 shows up to 3% worse performance than the method according to the invention when the arrival rate is 20 requests / minute, but worse by 21% when the arrival rate is 30 requests / minute. This can be explained as follows. In other words, when the arrival rate is 20 requests / minute, the buffer resource is not a problem in most cases, and thus CV3 shows only slightly worse performance than the method according to the present invention. However, as arrival speed increases, the server needs more buffer space to accommodate more customers, but this effect is more pronounced for the CV3 scheme, resulting in increased performance differences.

According to the present invention, by proposing an adaptive data retrieval method that dynamically changes the round length, it is possible to achieve load sharing between clusters without moving or copying data.

Claims (2)

The transfer rate is tr, the disk seek time is T _s , each constant, the buffer size is B, The time is divided into equally sized periods called rounds, and round-based scheduling is used where each authorized customer is served once in each round, The video stream V _i is divided into a finite number of consecutive segments, each segment being subdivided into NS sub-segments (where each sub-segment has a length BR of data retrieved during the basic round). Size), each segment is stored on disks in a round-robin fashion, and the data playback speed is dr _i , the unit of which is bits / sec, Consisting of C clusters, each cluster consisting of Q homogeneous disks, dividing the customers into C customer groups (CG ₁ , ..., CG _C ), and then the customers in group CG _k from the corresponding cluster k Configured to receive streams, and

In the clustered video server requesting this video stream V _i , When the disk bandwidth utilization DS _k (j) and the buffer utilization BS _k (j) for the cluster k when the round length is fr _j are given by Equations 1 and 2, respectively, <Equation 1>

(here,

) <Equation 2>

A method of adaptively retrieving data for load sharing in a clustered video server, (1) calculating the values of DS _k (j) and BS _k (j) for all clusters k (k = 1, ..., C) and maintaining their respective sets; (2) Disk utilization DS _m when the selection parameter SL _m is ND indicating that the SL _m th element of the set FS containing the possible round lengths in ascending order, fr _SLm is selected as the round length for cluster _m Obtaining a set SM of (ND) (m = 1, ..., C); (3) If a customer requests a video stream stored in cluster p, assume that a new customer will be authorized j = 1,... Recalculating the utilizations DS _p (j) and BS _p (j) for the cluster p for ND, where ND is the number of divisors for the NS; (4) determining whether the selection parameter SL _p for the cluster p is SL _p ≤ ND-1 to determine whether the round length can be extended for the cluster p; (5) Checking whether the disk bandwidth utilization DS _p (SL _p + 1) satisfies the following Equation 3 when increasing the selection parameter for the cluster p to determine whether it is possible to extend the round length without jitter. Doing; <Equation 3>

Where ε is the customer belonging to customer group CG _p

Maximum service time for) (6) If it is determined in step (4) to (5) that the round length is expandable without jitter, whether the buffer utilization in the case of extending the round length for the cluster p is smaller than 1 Checking by to determine whether there is sufficient buffer space; <Equation 4>

(7) if it is determined in step (6) that the buffer space is sufficient, increasing SL _p to SL _p + 1 to expand the round length; (8) selecting element DS _l (ND) having the smallest value in SM and removing it from SM; (9) checking whether the disk bandwidth utilization DS _l (ND-1), which is a case where the round length is reduced for the cluster l, is less than or equal to 1 to determine whether the disk bandwidth utilization is satisfied; And (10) If it is determined in step (9) that DS _l (ND-1) ≤ 1, SL _{l is set} to ND-1 to reduce the round length, otherwise step (8) to (8) while SM is not empty Repeat step (9) How to include. The method of claim 1, (11) if the customer closes the video stream stored in cluster k, recalculating DS _k (j) and BS _k (j) (j = 1, ..., ND); (12) determining whether the following Equation 5 is satisfied; <Equation 5>

(SL _k = 2,…, ND) (13) if it is determined in the step (12) that the equation (5) is satisfied, determining whether the following equation (6) is satisfied to determine whether reducing the round length would violate the disk bandwidth constraint; And <Equation 6>

(14) if it is determined in step (13) that the disk bandwidth constraint is not violated, reducing SL _k to SL _k -1 to reduce the round length of cluster k; How to include more.

Images