KR100806627B1 - A method for managing a video server including multi-speed disks - Google Patents

Abstract

A method for managing a video server including multi-speed disks is provided to reduce power consumption of the video server effectively without a jitter by using an energy-recognition data perfecting scheme for extending a low speed operation time of the multi-speed disks. A current speed level is compared with a selectable maximum speed level when a customer requests a video stream. It is checked whether first and second conditions are satisfied for the current speed level and the number of currently authenticated users when the current speed level is equal to a selectable maximum speed level. The customer is authenticated when the first and second conditions are satisfied, and is refused when the first and second condition are not satisfied. It is checked whether pre-fetching is advanced when the current speed level is lower than the selectable maximum speed. It is checked whether third, fourth, and fifth conditions are satisfied for the current speed level and the number of authenticated users when pre-fetching is not advanced. The customer is authenticated and pre-fetching for increasing a speed level is started for the data in a predetermined range when third, fourth, and fifth conditions are satisfied at the same time. The speed level is kept when the third, fourth, and fifth conditions are not satisfied.

Description

A METHOD FOR MANAGING A VIDEO SERVER INCLUDING MULTI-SPEED DISKS}

1 illustrates data retrieval during a prefetching window.

,

)의 서로 다른 값에 대하여, R1과 비교한 에너지 소비를 나타내는 도면으로서, 사용가능한 스피드의 개수를 증가시키는 것이 에너지의 소비를 감소시킨다는 것을 나타내는 도면.2 is (

,

A diagram showing energy consumption compared to R1 for different values of), wherein increasing the number of available speeds reduces the consumption of energy.

,

)의 서로 다른 값에 대하여, pf=2인 경우와 비교한 에너지 소비를 나타내는 도면으로서, 프리페칭 윈도우 크기가 클수록 에너지 소비를 감소시킨다는 것을 나타내는 도면.3 is (

,

A diagram showing the energy consumption compared to the case where pf = 2, for different values of), indicating that the larger the prefetching window size, the lower the energy consumption.

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

110: period of change (between speed levels)

120: prefetching window

130: prefetch A

140: prefetch B

150: Prefetch C

The present invention relates to a method for managing a video server comprising a multi-speed disc, and more particularly to a multi-speed disc using a new energy-aware data prefetching scheme that extends the length of time that the discs sustain at low speeds. The present invention relates to a method for effectively managing an included video server without jitter.

As the Internet supports more and more video applications such as digital libraries, education on demand, distance learning, and on-demand movies, the importance of video servers is becoming increasingly important. In a video server, video data is housed in storage and delivered in real-time to customers as requests from the customers occur.

Due to the increasing demand for data, server energy consumption is an important issue. Recently Energy User News reports that typical service providers currently require 150-200 W / ft ² of power and in the near future will require 200-300 W / ft ² of power. This growing demand for electricity can cause serious economic problems for service providers, for example, the 15 MW needed for a medium-sized 30,000 ft ² data center would cost $ 13,000,000 annually.

Another problem that can arise from higher power consumption is that of heat. It is reported that the failure rate can be doubled if the disk drive is operated at 15 ° C higher than standby. To address this, cooling systems for high thermal densities are enormously expensive, and the cooling system itself adds significantly to the power costs of the data center. Furthermore, much energy consumption makes it difficult to scale servers and negatively impact the environment. Of the components of the server, storage is one of the most energy consuming components. According to recent reports, storage devices consume about 27% of total power. It has also been shown that the energy consumed by the disk array may exceed the energy consumed in the rest of the system, depending on the size of the array. This problem can be exacerbated by the availability of higher speed disks that require more power.

One way to reduce the power consumption of the disk is to reduce the speed at which the disk spins. Gurumurthi et al. Have proposed a dynamic multi-speed disc model that can change the speed during rotation by using these characteristics. Using multi-speed disks like this can dramatically reduce power consumption even under concentrated server workloads. Currently multi-speed discs with a continuous range of speeds are not possible, but there are multi-speed discs that support a set of discrete speeds.

To save energy, the server needs to keep the disk in slow mode for as long as possible, but changing the speed requires considerable time. For example, in the dynamic multi-speed disk model, changing the speed takes 6.9 seconds. During the change period, the customer's request for data cannot be processed, which may cause distortion or pause of playback due to a violation of the time constraints of the video data. This phenomenon is called jitter. To ensure jitter-free playback, the server needs to prefetch data to play during the change. However, detailed analysis is required to determine the resources required for prefetching. Moreover, a significant portion of these resources must be reserved for prefetching, which usually results in under-utilization of the resources.

The operation of the disk can be divided into four phases: seek, active, idle and standby. In the search state, the arm moves across the platter and stops on the desired cylinder. In the active state, the disc actually reads or writes data. In the idle state, the disc is spinning but there is no customer request for service. In the standby state, the disc stops spinning completely. While the disk in standby consumes much less energy than the disk in the rest of the state, it needs to be rotated back to full speed to service the customer's request for data, which requires significant energy and Incurs a time penalty. Therefore, it makes sense to put the disk on standby only if the energy saved by putting the disk on standby is greater than the energy needed to rotate the disk again.

Typical power management for a single-speed disk involves two steps. The first step is to detect a suitable idle period, and the second step is to slow down the disk to standby. This is an efficient energy-saving technique in laptop computers where long idle periods can be expected and the disks have relatively low latency for spindown and spinup. However, large amounts of activity on enterprise servers require idle time between disks that is too short to justify the cost of spinning up and spinning down disks. As a result, the potential for energy saving is very limited.

Multi-speed discs are already in circulation, but only a few methods of changing the rotational speed of the disc have been studied. These approaches are mostly based on monitoring the length of the disk request queue and changing the speed of the disk so that the response time for the requests remains within a predetermined range. However, such power saving methods are not suitable for video server applications in that changing the speed in the video server requires considerable time, which may cause long stops in video playback. In addition, since video data generally requires a high data transfer rate, there is a problem that the disk has little chance of entering a low speed mode.

The present invention is derived from the above problem recognition, and can effectively reduce power without jitter in video servers including multi-speed discs by using a new data prefetching scheme that extends the length of time the discs continue to be slow. It is an object of the present invention to propose a video server management method.

According to a feature of the present invention for achieving the above object, a video server management method is as follows.

ND disks, each disk having NR speed levels (assuming that the rotational speed at speed level k is R (k), R (1) <… <R (NR)), In a video server where the change time between speed levels is T _o and the workload is distributed evenly among the disks,

이고, 프리페칭 이후의 고객들 수에서의 변화를 나타내는 파라미터는 Q이며(프리페칭이 시작되는 경우 Q가 0으로 설정됨),Prefetching for data to be used during periods of speed level change, the number of rounds during which prefetching is allowed, the size of the prefetching window is pf, and the server can accommodate additional customers during prefetching The number of

, The parameter representing the change in the number of customers after prefetching is Q (Q is set to 0 when prefetching starts),

The disk head scans back and forth along the surface of the disk, and SCAN scheduling is used where the block is retrieved when the head passes a block requested by the customer.

Round-based scheduling is used, where time is divided into equal-sized periods called rounds, each authorized customer being served once in each round, and each The length of the round is R,

The display speed required by all video streams is dr (bits / sec), the disk seek time for each disk is T _s , the buffer size is B,

The data transfer rate of a disk operating at speed level k is tr (k), the rotational latency is L (k),

Assuming no prefetching, the disk bandwidth utilization DU _no (NA, k) and buffer utilization BU _no (NA) when the speed level is k and the number of customers is NA names are represented by the following equations 1 and 2, respectively. Given by

로서 주어짐)(Where ST _no (NA, k) is the service time for one disk when the speed level is k and the number of customers is NA in case of ignoring the additional bandwidth for prefetching,

Given as

Assuming prefetching, the disk bandwidth utilization DU _pr (NA, k) and buffer utilization BU _pr (NA) when the speed level is k and the number of customers is NA names are given by the following equations (3) and (4), respectively. Is given by

로서 주어지며, 여기서 SF_pr(k)는 속도 레벨 k에서 비디오 스트림 하나를 프리페칭하기 위해 요구되는 서비스 시간으로서,

로서 주어짐)(Where ST _pr (NA, k) is the service time for one disk when the speed level is k and the number of customers is NA in case of considering the additional bandwidth for prefetching,

Where SF _pr (k) is the service time required to prefetch one video stream at rate level k,

Given as

A method of managing the video server,

(1) when the customer requests a video stream, comparing the currently selected speed level CR with a selectable maximum speed level NR;

(조건 1) 및

(조건 2)를 만족하는지 확인하여, 상기 조건 1 및 2를 동시에 만족하는 경우 상기 고객을 인가하고, 만족하지 못하는 경우 상기 고객을 거절하는 단계;(2) As a result of the comparison of step (1), when CR = NR, for the currently selected speed level CR and the number of currently authorized customers CA,

(Condition 1) and

Confirming whether (condition 2) is satisfied and approving the customer if the conditions 1 and 2 are satisfied at the same time, and rejecting the customer if the condition is not satisfied;

(3) confirming whether prefetching is currently in progress when CR <NR as a result of the comparison in step (1);

(조건 3),

(조건 4) 및

(조건 5)를 만족하는지 확인하는 단계;(4) As a result of the checking in step (3), if prefetching is not currently in progress, for the currently selected speed level CR and the number of currently authorized customers CA,

(Condition 3),

(Condition 4) and

Checking whether (condition 5) is satisfied;

~

(여기서, ro는 현재 라운드임) 범위의 데이터에 대하여 속도 레벨 증가를 위한 프리페칭을 시작하며, Q를 0으로 설정하는 단계;(5) If the check result of step (4) indicates that the conditions 3 to 5 are satisfied at the same time, the customer is authorized,

To

Starting prefetching for speed level increase over data in the range, where ro is the current round, setting Q to 0;

(조건 1) 및

(조건 2)를 만족하는지 확인하여, 상기 조건 1 및 2를 동시에 만족하는 경우 상기 고객을 인가하고, 만족하지 못하는 경우 상기 고객을 거절하는 단계; 및(6) If the result of the check in step (4) is not satisfied, the speed level is maintained, and at the same time, for the currently selected speed level CR and the number of currently authorized customers CA,

(Condition 1) and

Confirming whether (condition 2) is satisfied and approving the customer if the conditions 1 and 2 are satisfied at the same time, and rejecting the customer if the condition is not satisfied; And

이면 상기 고객을 인가하고, 아니면 속도 변화 완료시까지 상기 고객의 인가를 지연하는 단계를 포함하는 것을 그 특징으로 한다.(7) As a result of confirming the above step (3), if prefetching is currently in progress, after increasing the value of Q by 1,

And approving the customer, or delaying the authorization of the customer until the speed change is completed.

According to another aspect of the present invention, a video server management method includes

(8) if the customer closes the video stream, checking whether prefetching is currently in progress;

(조건 6) 및

(조건 5)를 만족하는지 확인하는 단계;(9) As a result of the checking in step (8), if prefetching is not currently in progress, for the currently selected speed level CR and the number of currently authorized customers CA,

(Condition 6) and

Checking whether (condition 5) is satisfied;

~

(여기서, ro는 현재 라운드임) 범위의 데이터에 대하여 속도 레벨 감소를 위한 프리페칭을 시작하며, Q를 0으로 설정하는 단계;(10) when the result of checking (9) above satisfies the conditions 5 and 6 simultaneously,

To

Starting prefetching for speed level reduction for data in the range (where ro is the current round), setting Q to 0;

(11) maintaining the speed level if it is not satisfied as a result of the checking in step (9); And

이면 프리페칭을 취소하는 단계를 더 포함하는 것을 그 특징으로 한다.(12) As a result of confirming the above step (8), if prefetching is currently in progress, after decreasing the value of Q by 1,

And canceling prefetching.

A video server management method according to another aspect of the present invention,

It is characterized by the fact that the display speed dr (bits / sec) required by each video stream is not a constant.

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

1. System and multi- speed  Disk model

1.1 System Model

To support multiple customers, round-based scheduling is used for data retrieval. That is, after time is divided into equally sized periods called 'rounds', each authorized customer is served once in each round. We will assume that the amount of data retrieved during a round corresponds to one disk block. This hypothesis provides good performance in video data retrieval because only one data retrieval is required to read one video during each round. In general, video data has real-time requirements. That is, the video frame can be used only when the video frame is retrieved before the playback time. To ensure continuous playback, the data required for the current playback must have been read from the disc in the previous round. On the other hand, data to be reproduced in the next round is read during the current round. This means that in order to achieve jitter-free playback, the service time (ie the total time spent searching for all streams for a round) must not exceed the duration of one round.

Current commercial multi-speed discs have only two speeds. However, since there seems to be no fundamental obstacle to supporting more speeds, we will consider a disk with multiple discrete speeds. To support the high bandwidth and large storage requirements of video data, servers typically use disk arrays. In the proposed model, the disk array consists of the same multi-speed disks. To reduce the search overhead, SCAN scheduling is used where the disk head scans back and forth along the surface of the disk, and when the head passes a block requested by the customer.

1.2 multi- speed  disk Modeling

Disks have several components that contribute to overall power consumption. This includes a spindle motor that rotates the platter, an actuator that moves the head, electrical components and other circuitry involved in data transmission. In general, the spindle motor is responsible for about 80% of the total idle-mode power consumption, so the spindle motor is particularly important for modeling disk power consumption. Assume a disk has NR speed levels. Let R (k) be the rotational speed at the speed level k. R (1) < Assume <R (NR). The power P _m (k) required by the spindle motor generally has a linear or quadratic relationship with the speed.

If the speed level is k, then P _s (k) is the power required during the search state, P _a (k) is the power required during the active state, and P _i (k) is the power consumed during the idle state. . P _b is the power consumption during the standby state. In the idle state, the power P _o consumed by the other components of the disk except the spindle motor can be considered independent of the platter speed. Thus P _i (k) = P _m (k) + P _o . While active, the data channel is actively transferring bits between electronics and platters. This includes the movement of the read / write head, which requires additional power P _h . Since the power required to move the read / write head is independent of the rotational speed, P _a (k) can be expressed as P _i (k) + P _h . During the search state, the actuator needs to move the arm onto the desired cylinder, which requires as much power as P _g than the idle state. As a result, P _s (k) = P _i (k) + P _g .

Multi-speed discs require a significant amount of time to change speed because they can only change speed in standby. Since data cannot be read from the disk during the change period, the server needs to prefetch data to be consumed during this period. It will be assumed that the speed change is allowed only between two adjacent speed levels. Let T _{o be the} time of change between the two speed levels. For simplicity, assume T _o is a multiple of round length R.

2. Energy Awareness Data Prefetching

2.1 at different speed levels resource  Use

Let tr (k) be the data transfer rate (bits per second) of a disk operating at speed level k (k = 1, ..., NR). We will use a typical search time model that requires a constant search time for one search operation on adjacent data. Note that the search time is independent of the rotational speed, so that the constant Ts is used to indicate the search time. Since the rotational delay varies with the speed level, L (k) is used to indicate the rotational latency of the disc at speed level k. For ease of explanation, it will be assumed that all video streams require the same display rate of dr bits / second. However, the approach according to the present application can be easily adjusted to work with different display speeds.

의 판독 시간을 발생시킨다. 결과적으로, 하나의 비디오 스트림을 서비스하는 것은

의 서비스 시간을 증가시킨다. 프리페칭을 위한 부수적인 디스크 대역폭을 무시하면, 속도 레벨 k에서의 디스크 하나에 대한 서비스 타임

은 다음 수학식 5와 같이 표현될 수 있다.Let's see how the disk bandwidth requirements of the video depend on the speed level k and the number of authorized customers NA. For a disc running at speed level k, once retrieving the video stream is equivalent to the retrieval and rotation delay overhead of T _s + L (k).

Generates a reading time of. As a result, servicing a single video stream

Increase the service time. Ignoring the additional disk bandwidth for prefetching, service time for one disk at speed level k

May be expressed as Equation 5 below.

는 다음 수학식 6과 같이 계산될 수 있다.Disk bandwidth utilization is defined as the ratio of total service time to round length R. Suppose there are ND disks in the server. In addition, for ease of explanation, it is assumed that the workload is distributed evenly among the disks. By delaying the authorization of customers, the load can be easily balanced. If you ignore prefetching, disk bandwidth utilization for NA customers

May be calculated as in Equation 6 below.

If the server does not have enough disk bandwidth to support other streams, the speed level can be increased to accommodate new customers. On the other hand, if disk bandwidth utilization is low enough, the speed level can be reduced to save power. In either case, the server must prefetch data to be consumed during the rate change period. Thus we introduce the concept of a prefetching window, which is the number of rounds while prefetching is allowed. 1 is a diagram illustrating data retrieval during a prefetching window. As shown in FIG. 1, if pf is the size of the prefetching window, the server starts prefetching data to be consumed for the pf round corresponding to the change period before the change actually begins. Since the disk reads the stop command during the speed change, the server needs to prefetch enough data for T _o seconds. It is also necessary to prefetch data to be consumed during the round immediately after the change period to ensure jitter-free playback. Therefore, the entire T _o + R has a second data to be prefetched for each video stream.

의 판독 시간을 발생시킨다. j개의 라운드에 대한 데이터를 판독하는 것은, j번의 검색 및 회전 지연 오버헤드를 발생시킨다.

개의 라운드에 대한 데이터가 프리페치 될 필요가 있으므로, 하나의 스트림을 프리페칭하기 위해 요구되는 서비스 시간

은 다음 수학식 7과 같이 계산될 수 있다.Since enough data for T _o + R seconds needs to be prefetched, the prefetching operation

Generates a reading time of. Reading data for j rounds incurs j search and rotation delay overhead.

Service time required to prefetch one stream, because data for three rounds needs to be prefetched

May be calculated as in Equation 7 below.

은, 다음 수학식 8과 같이 계산될 수 있다.Since the server prefetches data through the prefetch window, it can distribute the prefetch operation among the windows. As can be seen in FIG. 1, the data for stream A may be prefetched in the i-th round, the data for stream B may be prefetched in the (i + 1) -th round, and in stream C The data for can be prefetched in the (i + 2) -th round. Service time required to prefetch data for NA customers

May be calculated as in Equation 8.

Since disk bandwidth needs to be prepared for prefetching, and there are ND disks, the following equation 9 for disk bandwidth utilization considering prefetching can now be obtained.

Now let's see how the buffer requirement for one video depends on the number NA of authorized customers. Let B be the total buffer size. Double-buffering is used for SCAN scheduling. Accordingly, if the buffer requirement for prefetching is ignored, the buffer utilization may be expressed as in Equation 10 below.

을 수학식 10에 추가함으로써, 프리페칭을 고려한 버퍼 이용도

을 다음 수학식 11과 같이 얻을 수 있다.To prepare for a change in speed level, an additional buffer space of T _o x dr is required to account for the change period and an additional buffer space of R x dr is required to account for the round immediately after the change period. therefore

Is added to Equation 10, buffer utilization considering prefetching

Can be obtained as in Equation 11 below.

2.2 Speed Level Changes

1) Change in speed level without jitter

As the server load changes, the speed level can change. By default, if the system load is low, a low speed level will probably be chosen. On the other hand, high speed levels are more suitable for overload. The difficulty here is determining exactly when to change the speed level without jitter. In this application, we use a threshold-based approach in which a change in speed level is triggered by a change in disk bandwidth utilization.

명의 추가 고객을 수용하기에 충분한 리소스를 예비한다. 일반적으로, 이 값은 고객 도착의 패턴에 대한 계속적인 측정에 의해 결정된다.Note that the number of customers may increase during the prefetching window. To reflect this, the server will

Reserve enough resources to accommodate two additional customers. In general, this value is determined by a continuous measure of the pattern of customer arrival.

와

가 요구된다. 본 출원에 따른 접근법에서는, 서버는

가 1의 임계값을 초과하기 바로 전에 속도 레벨을 증가시킨다. 보다 상세하게는, 만약 사용가능한 디스크 대역폭 이용도가 한 명의 고객에 대한 디스크 대역폭 이용도를 표시하는

보다 작으면, 서버는 속도 레벨을 증가시킨다. 따라서 만약

이고

이면, 속도 레벨이 증가된다. 역으로, 만약

이면, 속도 레벨은 감소된다.Assume the currently selected speed level for the disk is CR (CR = 1, ..., NR), and CA is the number of customers currently authorized. To increase the jitter-free speed level,

Wow

Is required. In the approach according to the present application, the server is

Increase the speed level just before it exceeds the threshold of 1. More specifically, if the available disk bandwidth utilization indicates the disk bandwidth utilization for one customer

If less, the server increases the speed level. So if

ego

If so, the speed level is increased. Conversely, if

If so, the speed level is reduced.

의 값을 감소시키고, 프리페칭을 위해 준비되는 디스크 대역폭을 감소시키며, 따라서 디스크들이 저속 모드에서 지속하는 시간의 길이를 연장할 더 많은 기회를 제공한다는 것을 알 수 있다. 그러나 pf의 값이 증가하면, 미래의 pf 라운드에서의 고객의 수를 예측하기가 더 어렵게 된다. 이것은 S1 및 S2와 같은 다음 2가지 원하지 않는 상황을 초래할 수 있다.From Equation 8, increasing the window size pf,

It can be seen that reducing the value of and reducing the disk bandwidth prepared for prefetching, thus providing more opportunities for the disks to extend the length of time they last in slow mode. However, as the value of pf increases, it becomes more difficult to predict the number of customers in future pf rounds. This can lead to the following two undesirable situations such as S1 and S2.

의 조건임에도 불구하고 속도 레벨이 증가될 수 있다.S1:

The speed level can be increased despite the condition of.

의 조건임에도 불구하고 속도 레벨이 감소될 수 있다.S2:

The speed level can be reduced despite the condition of.

In situation S1, the speed level is unnecessarily increased, which may affect the disk life. Also, if situation S2 occurs, the next increase in speed level can result in jitter. To avoid both situations, we will now specifically describe the rate change algorithm applied when a customer requests or closes a video stream.

2) If customer requests video stream

이고

인지를 체크한다. 만약 이 조건이 만족되면, 버퍼 조건

을 체크한다. 만약 이 조건 또한 만족되면, 서버는 새 고객을 인가하고 라운드

와

사이에 검색될 데이터를 프리페칭 하기 시작한다(여기서 ro는 현재의 라운드이다). 그렇지 않으면, 프리페칭이 지터를 발생시킬 수 있기 때문에 속도 레벨은 증가될 수 없다.In this case, the server needs to increase the speed level to accommodate new customers. If CR <NR and prefetching is not ready, the server needs to determine if the speed level can be increased. For this purpose, the server

ego

Check for acknowledgment. If this condition is met, the buffer condition

Check If this condition is also satisfied, the server authorizes the new customer and rounds

Wow

Start prefetching the data to be retrieved between (where ro is the current round). Otherwise, the speed level cannot be increased because prefetching can cause jitter.

이면, 서버가

명까지의 고객에 대한 추가적인 리소스를 예비하였기 때문에 서버는 새 고객을 받아들인다. 그렇지 않으면, 속도 변화가 완료될 때까지 서버는 고객의 인가를 지연시킨다. 고객의 수는 프리페칭 후에 감소될 수도 있다는 점에 주목하자. 이 경우에, 프리페칭을 하는 것은 상황 S1을 초래할 수 있기 때문에 서버는 프리페칭을 취소시킬 필요가 있다. 따라서

이면, 서버는 프리페칭을 취소하고 모든 프리페치 된 데이터를 버린다.To prevent the situation S1, the server maintains a parameter Q indicating a change in the number of customers after prefetching. This value is set to zero when prefetching is started. If prefetching is being prepared and a new customer requests a video stream, the server increases the value of Q.

If the server

The server accepts new customers because it has reserved additional resources for up to three customers. Otherwise, the server delays the authorization of the customer until the speed change is complete. Note that the number of customers may decrease after prefetching. In this case, the server needs to cancel the prefetching because prefetching may result in situation S1. therefore

If so, the server cancels prefetching and discards all prefetched data.

이면, 서버는 현재의 속도 레벨을 유지한다. 만약 프리페칭을 위한 충분한 버퍼 공간이 없다면(

), 속도 레벨을 증가시키는 것은 지터를 발생시킬 수 있기 때문에 서버는 속도 레벨을 증가시키려는 시도를 그만둔다. 만약 CR = NR이면, 서버는 속도 레벨을 증가시킬 수 없는 동일한 위치에 있다. 이들 경우에, 속도 변화가 없을 것이기 때문에 서버는 프리페칭을 무시하는 다른 인가 조건을 적용한다. 이 조건은

이다. 만약 이 조건이 위반되면, 새 고객은 거절된다. 그렇지 않으면, 서버는 버퍼 제약조건

을 체크한다. 만약 이 조건이 만족되면, 새 고객이 인가된다. 그렇지 않으면, 새 고객은 거절된다.if

If so, the server maintains the current speed level. If there is not enough buffer space for prefetching (

Increasing the speed level can cause jitter, so the server stops trying to increase the speed level. If CR = NR, the server is in the same position that cannot increase the speed level. In these cases, since there will be no speed change, the server applies another authorization condition that ignores prefetching. This condition

to be. If this condition is violated, the new customer will be rejected. Otherwise, the server will issue a buffer constraint

Check If this condition is met, the new customer is authorized. Otherwise, the new customer is rejected.

3) If the customer closes the video stream

인지를 체크한다. 만약 이 조건이 만족되면, 서버는

인지를 체크한다. 만약 이 조건 또한 만족되면, 속도 레벨에서의 감소가 레벨의 다음번 증가에서 지터를 발생시키지 않을 것이며, 따라서 감소가 허가된다. 그렇지 않으면, 현재 속도 레벨이 유지된다. 허가된 고객들의 수가 프리페칭 후에 증가될 수 있다는 점에 주목하자. 이 경우에, 프리페칭은 상황 S2를 초래할 수 있기 때문에, 서버는 프리페칭을 취소할 필요가 있다.In this case, the server can save energy by reducing the speed level. For this purpose, the server

Check for acknowledgment. If this condition is met, the server

Check for acknowledgment. If this condition is also satisfied, a decrease in the speed level will not cause jitter at the next increase in the level, so a decrease is permitted. Otherwise, the current speed level is maintained. Note that the number of authorized customers can be increased after prefetching. In this case, since prefetching can result in situation S2, the server needs to cancel the prefetching.

3. Experimental Results

3.1 Set up experimental conditions Experimentalal Setup )

To evaluate the effectiveness of the proposed method, we simulated a server with six disks with a maximum data transfer rate of 61.4 MB / s and a typical search time of 11.2 ms. Up to five speed levels between 28800 and 72000 RPM (Rotation Per Minute) were considered. P _i (k) was calculated using a linear model. P _h and P _g are assumed to be 3.3 W. Detailed parameters of each disk are shown in Table 1 below.

Rotational Speed (RPM) 28800 39600 50400 61200 72000 Baud rate (MB / s) 24.56 33.77 42.98 52.19 61.4 Rotation delay (ms) 5 3.6 2.86 2.35 2 Idle power (W) 4.08 5.61 7.14 8.67 10.2 Active / Search Power (W) 7.38 8.91 10.41 11.97 13.5

인 Zipf 분포를 따른다.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. The round length is 3.5 seconds. The speed change time is 7 seconds and the change energy is 152J. Access probability

Follow the Zipf distribution.

초로부터

초까지 변하며, 다음 10시간 동안은

초로부터

초까지 변하는데, 여기서

초의 내부-도착 시간은 피크 부하에서의 한 시간에 해당한다.Server load generally follows a diurnal cycle. To allow this, the energy consumption was examined for more than 20 hours, and the rate at which requests reached on an hourly basis was changed as follows. That is, the inter-arrival time is the first 10 hours

From the beginning

Changes to seconds, and for the next 10 hours

From the beginning

Changes to seconds, where

The inner-arrival time of seconds corresponds to one hour at peak load.

3.2 Effect of Speed Setting Effects of Speed Sets )

,

)의 서로 다른 값에 대하여, R1과 비교한 에너지 소비를 나타내는 도면이다. 단지 하나의 속도만을 갖는 R1은 항상 가장 나쁜 성능을 보인다. 따라서 예컨대, R5는 12%와 45% 사이만큼, R3은 10%와 42% 사이만큼, R2는 3%와 37% 사이만큼 R1보다 적은 에너지를 소비한다. 이용도가 상대적으로 낮은 경우(예컨대,

=11,

=20), 속도 집합들 사이의 성능 차이는 증가한다. 모든 집합은 72000 RPM이라는 동일한 최대 속도를 지원하기 때문에 허용될 수 있는 고객들의 수가 각각의 경우에 동일하다는 점에 주목하자. 이들 결과는 멀티-스피드 디스크가 비디오 서버 작업부하 하에서 상당한 에너지 절약을 제공한다는 것을 알려준다. 또한, 도 2는 속도의 개수를 증가시키는 것이 에너지 소비를 감소시킨다는 것을 도시한다. 이것은, 사용가능한 보다 많은 속도를 가짐으로써 서버가 저속을 이용할 보다 많은 기회를 가질 수 있기 때문이다.First, we examined how energy consumption varies with the number of available speeds. The values were R1 = 72000, R2 = 28800, 72000, R3 = 28800, 50400, 72000, R5 = 28800, 39600, 50400, 61200, 72000. Four disks with different speed sets were considered. The value of pf is assumed to be 15 seconds. 2 is (

,

It is a figure which shows the energy consumption compared with R1 about the different value of (). R1 with only one speed always shows the worst performance. Thus, for example, R5 consumes less energy than R1 by between 12% and 45%, R3 by between 10% and 42%, and R2 by between 3% and 37%. Relatively low utilization (e.g.,

= 11,

20), the performance difference between the speed sets increases. Note that all sets support the same maximum speed of 72000 RPM, so the number of customers that can be allowed is the same in each case. These results indicate that multi-speed discs provide significant energy savings under video server workloads. 2 also shows that increasing the number of speeds reduces energy consumption. This is because having more speed available allows the server to have more opportunities to take advantage of the lower speed.

3.3 Prefetching window  Effect of size

,

)의 서로 다른 값에 대하여, pf=2인 경우와 비교한 에너지 소비를 나타내는 도면이다. 도 3으로부터, 프리페칭 윈도우 크기가 클수록 에너지 소비를 감소시킨다는 것을 알 수 있다. 예를 들어, 만약 pf=14이면, 디스크는 pf=2 경우보다 9%와 36% 사이만큼 작은 에너지를 사용한다. 이것은 다음과 같이 설명될 수 있다. pf를 증가시키는 것은, 지터가 없는 속도 증가를 위한 디스크 대역폭 요구조건을 완화시켜 주며, 따라서 pf의 높은 값에 대해 변하지 않고 유지되어야 할 조건에서 pf가 낮은 경우 속도 레벨이 증가될 수 있기 때문이다. 본 발명에서 제안된 방법은, 프리페칭을 취소시킴으로써 윈도우 크기가 증가하면 발생할 수 있는 상황 S1 및 S2를 해결하는데, 이것은 윈도우 크기를 증가시킬 수 있는 보다 많은 기회를 제공한다.When using velocity set R5, we investigated how the prefetching window size affects power consumption. 3 is (

,

It is a figure which shows energy consumption compared with the case where pf = 2 with respect to the different value of (). It can be seen from FIG. 3 that the larger the prefetching window size, the lower the energy consumption. For example, if pf = 14, the disk uses less energy between 9% and 36% than pf = 2. This can be explained as follows. Increasing pf relaxes the disk bandwidth requirement for jitter-free speed growth, so the speed level can be increased if pf is low under conditions that must remain unchanged for high values of pf. The method proposed in the present invention addresses situations Si and S2 that may occur if the window size increases by canceling prefetching, which provides more opportunities to increase the window size.

와 같이 계산할 수 있는데, 이 값은 pf가 증가함에 따라 감소한다. 결과적으로, 프리페칭을 위해 요구되는 서비스 시간에서의 차이가 감소되어, 도 3에 도시된 것과 같은 결과를 초래한다. pf를 증가시키는 것이 항상 에너지 소비를 줄이는 것은 아니라는 점에 주목할 만하다. 예를 들어,

=2이고

=20인 경우, pf=50으로 설정하는 것은 pf=14보다 2% 저하된 성능을 보인다. 이것은 중단된 프리페칭 동작상에 작용된 전력 때문이다. 만약 pf>14이면, 에너지 소비에서의 차이가 2%보다 적다는 것을 관찰할 수 있다. 이것은 윈도우 크기를 최대한 증가시키는 것 대신에, 타당한 값의 일부 범위가 주어져야만 한다는 것을 의미한다.In addition, it can be observed from FIG. 3 that as the pf increases, the difference in energy consumption between two adjacent window sizes becomes smaller. For example, the difference between pf = 12 and pf = 14 is between 0.02% and 1%, while the difference between pf = 2 and pf = 4 is between 2% and 22%. From Equation 8, the difference in service time of two adjacent window sizes between pf and pf + 1

It can be calculated as: This value decreases as pf increases. As a result, the difference in service time required for prefetching is reduced, resulting in a result as shown in FIG. It is noteworthy that increasing pf does not always reduce energy consumption. E.g,

= 2

In the case of = 20, setting pf = 50 is 2% lower than pf = 14. This is due to the power applied to the interrupted prefetching operation. If pf> 14, it can be observed that the difference in energy consumption is less than 2%. This means that instead of increasing the window size as much as possible, some range of valid values must be given.

The present invention described above may be variously modified or applied by those skilled in the art, and the scope of the technical idea according to the present invention should be defined by the following claims.

According to the video server management method of the present invention, it is possible to effectively reduce power without jitter in a video server including a multi-speed disc by using a new data prefetching scheme that extends the length of time that the discs sustain at low speed.

Claims (3)

ND disks, each disk having NR speed levels (assuming that the rotational speed at speed level k is R (k), R (1) <… <R (NR)), In a video server where the change time between speed levels is T _o and the workload is distributed evenly among the disks, Prefetching for data to be used during periods of speed level change, the number of rounds during which prefetching is allowed, the size of the prefetching window is pf, and the server can accommodate additional customers during prefetching The number of

, The parameter representing the change in the number of customers after prefetching is Q (Q is set to 0 when prefetching starts), The disk head scans back and forth along the surface of the disk, and SCAN scheduling is used where the block is retrieved when the head passes a block requested by the customer. Round-based scheduling is used, where time is divided into equal-sized periods called rounds, each authorized customer being served once in each round, and each The length of the round is R, The display speed required by all video streams is dr (bits / sec), the disk seek time for each disk is T _s , the buffer size is B, The data transfer rate of a disk operating at speed level k is tr (k), the rotational latency is L (k), Assuming no prefetching, the disk bandwidth utilization DU _no (NA, k) and buffer utilization BU _no (NA) when the speed level is k and the number of customers is NA names are represented by the following equations 1 and 2, respectively. Given by <Equation 1>

(Where ST _no (NA, k) is the service time for one disk when the speed level is k and the number of customers is NA in case of ignoring the additional bandwidth for prefetching,

Given as <Equation 2>

Assuming prefetching, the disk bandwidth utilization DU _pr (NA, k) and buffer utilization BU _pr (NA) when the speed level is k and the number of customers is NA names are given by the following equations (3) and (4), respectively. Is given by <Equation 3>

(Where ST _pr (NA, k) is the service time for one disk when the speed level is k and the number of customers is NA in case of considering the additional bandwidth for prefetching,

Where SF _pr (k) is the service time required to prefetch one video stream at rate level k,

Given as <Equation 4>

A method of managing the video server, (1) when the customer requests a video stream, comparing the currently selected speed level CR with a selectable maximum speed level NR; (2) As a result of the comparison of step (1), when CR = NR, for the currently selected speed level CR and the number of currently authorized customers CA,

(Condition 1) and

Confirming whether (condition 2) is satisfied and approving the customer if the conditions 1 and 2 are satisfied at the same time, and rejecting the customer if the condition is not satisfied; (3) confirming whether prefetching is currently in progress when CR <NR as a result of the comparison in step (1); (4) As a result of the checking in step (3), if prefetching is not currently in progress, for the currently selected speed level CR and the number of currently authorized customers CA,

(Condition 3),

(Condition 4) and

Checking whether (condition 5) is satisfied; (5) If the check result of step (4) indicates that the conditions 3 to 5 are satisfied at the same time, the customer is authorized,

To

Starting prefetching for speed level increase over data in the range, where ro is the current round, setting Q to 0; (6) If the result of the check in step (4) is not satisfied, the speed level is maintained, and at the same time, for the currently selected speed level CR and the number of currently authorized customers CA,

(Condition 1) and

Confirming whether (condition 2) is satisfied and approving the customer if the conditions 1 and 2 are satisfied at the same time, and rejecting the customer if the condition is not satisfied; And (7) As a result of confirming the above step (3), if prefetching is currently in progress, after increasing the value of Q by 1,

If the customer is authorized, or delaying the approval of the customer until the speed change is completed. How to include. The method of claim 1, (8) if the customer closes the video stream, checking whether prefetching is currently in progress; (9) As a result of the checking in step (8), if prefetching is not currently in progress, for the currently selected speed level CR and the number of currently authorized customers CA,

(Condition 6) and

Checking whether (condition 5) is satisfied; (10) when the result of checking (9) above satisfies the conditions 5 and 6 simultaneously,

To

Starting prefetching for speed level reduction for data in the range (where ro is the current round), setting Q to 0; (11) maintaining the speed level if it is not satisfied as a result of the checking in step (9); And (12) As a result of confirming the above step (8), if prefetching is currently in progress, after decreasing the value of Q by 1,

Step to cancel prefetching How to include more. The method according to claim 1 or 2, The display speed dr (bits / sec) required by each video stream is not a constant.

Images