CN103064730B - A kind of two-stage disk-scheduling method of facing cloud computing environment - Google Patents

Abstract

A two-level disk scheduling method oriented to a cloud computing environment belongs to the technical field of computer storage, and reduces the problem of uncoordinated two-level disk scheduling existing in the existing two-level disk scheduling method. The present invention includes an initialization step, a monitoring step, a prediction step and a decision step. The monitoring step monitors in real time the information reflecting the merging status of I/O requests, and provides the information to the decision-making step to determine whether the upstream scheduling is in the state of "over-merging" or "under-merging". The monitoring step and the decision-making step cooperate with each other to form a feedback control, which reduces the inconsistency of the two-level disk scheduling existing in the existing two-level disk scheduling method, and reduces the disk scheduling of the drive domain because of the "over-merging" or "under-merging" of the disk scheduling in the virtual machine. "And the adverse effects on I/O request merging are brought about, thereby improving the overall storage performance of the cloud computing system.

Description

A Two-Level Disk Scheduling Method for Cloud Computing Environment

technical field

The invention belongs to the technical field of computer storage, and in particular relates to a two-level disk scheduling method for a cloud computing environment.

Background technique

Cloud computing technology can enable multiple guest operating systems (Guest OS) to be deployed in different virtual machines (Virtual Machine) and run on a physical machine at the same time. This resource sharing mode can improve the reuse level of hardware resources, reduce system management costs, and improve the security of device access.

In order to prevent virtual machines from interfering with each other when accessing disks, especially when a virtual machine has an abnormality in disk access, the normal access of other virtual machines will not be affected, the existing cloud computing system adopts the Split Driver Model (Split Driver Model). ) or similar techniques for reliable disk access. As shown in Figure 1, the driver separation model sets the disk driver in a dedicated driver domain (Driver Domain), and the virtual machine does not directly access the disk device, but sends I/O request information through the front driver (Front Driver) To an I/O buffer (I/O Ring) and notify the driver domain through the message channel (EventChannel), the driver domain takes out the I/O request of the I/O buffer through the back-end driver (Back Driver) and sends it to the disk native driver for processing. After the I/O request is executed, the driver domain sends the I/O requested data back to the I/O buffer, and notifies the front-end driver of the virtual machine through the message channel again, and the virtual machine retrieves the I/O requested data through the front-end driver . This disk I/O request and execution mode is transparent to the operating system in the virtual machine, that is, the application program in the virtual machine will think that it directly accesses the disk. This drive separation model or similar technology objectively makes the disk I/O requests undergo two-level disk scheduling control, that is, the disk scheduling of the virtual machine operating system (upstream scheduling) and the disk scheduling in the driver domain (downstream scheduling). The degree of coordination of the two-level disk scheduling will seriously affect the performance of disk access.

In disk scheduling, I/O request coalescing is one of its optimization functions. Due to the mechanical operation mode of the magnetic head in the disk, the swing of the magnetic head will generate a large time overhead, which will affect the disk access performance, and I/O request merging can reduce the frequency of the magnetic head swing, thereby reducing the additional overhead caused by the head seek. time overhead. In addition, due to the reduction of I/O requests, the number of operating system and disk device command interactions will also be reduced, which also greatly reduces the protocol overhead of disk access (such as the ISCSI protocol). Therefore, I/O request coalescing is crucial in disk access. The necessary condition for I/O request merging is that the I/O requests participating in the merging must have adjacent Logical Block Addresses (Logic Block Address, LBA for short). As shown in Figure 2, when the initial LBA of the I/O request (Ra) is equal to the termination LBA of a certain I/O request (Rb) in the I/O request queue in the disk scheduling (the initial LBA plus the I/O request size), called Back Merge; and when the terminating LBA of an I/O request (Ra) is equal to the starting LBA of an I/O request (Rb) in the disk scheduling queue, it is called Front Merge (Front Merge). Merge). In addition, when the I/O request participating in the merging arrives at the disk scheduling, another I/O request participating in the merging must also be in the scheduling, otherwise, the I/O request merging still cannot be completed.

In the two-level disk scheduling architecture, I/O request merging is jointly completed by the disk scheduling in the virtual machine and the disk scheduling in the driver domain. As shown in Figure 3, the I/O request sent by the application deployed in the virtual machine must first be processed by the file system layer (including the virtual file system and the actual file system (such as Ext2 or Ext3, etc.) I/O requests of the same size, and then these I/O requests arrive at the upstream scheduling for request merging, and are sent to the driver domain, and the downstream scheduling further completes the I/O request merging. In this way, the cooperation degree of the two-level disk scheduling will directly affect the degree of I/O request merging. In the driver separation model of cloud computing, due to the limitation of the operating system on I/O request merging and the efficiency factor of virtualization, the degree of I/O request merging scheduled in the virtual machine needs to be limited. Existing virtualization software and operating systems generally adopt a unified parameter setting, and if modification is required, the operating system must be recompiled. This is not good for performance optimization of virtual I/O.

In the cloud computing environment, I/O request merging is usually completed by two-level disk scheduling in pipeline mode, so the downstream scheduling of the pipeline (disk scheduling in the driver domain) is an extension of its upstream scheduling (disk scheduling in the virtual machine).

Existing cloud computing systems generally use a fixed system configuration to constrain the degree of I/O request merging for upstream scheduling (that is, limit the number of I/O request merging). When the size of the I/O request or the number of merges exceeds the system limit, the two I/O requests that could have been merged will receive a merge abort signal, so that the merge operation cannot be scheduled at this level. If you want to modify the system configuration of the degree of I/O request consolidation, you must recompile the operating system and cloud computing system, which makes it even more difficult to adapt to the changes in the I/O request of the deployment program in the virtual machine, so that the "over-consolidation" of upstream scheduling and The "under-merge" problem affects the improvement of disk I/O performance.

"over-merging":

Assuming that the two I/O requests that can be merged are called the predecessor request Ra and the successor request Rb according to the order in which they arrive at the disk scheduling, the prerequisite for request merging is that when Rb arrives at the disk scheduling, Ra must still be in the scheduling, otherwise, The request merge cannot be completed. Although, due to the difference in the scheduling policy and the characteristics of the I/O request sent by the application, the time that the request stays in the disk scheduling is different, but, from the perspective of probability, the time difference between the predecessor request Ra and the successor request Rb has a great impact on the request merging Greater impact. Assuming that the maximum time limit for a request to stay in disk scheduling is 250ms, then if the time difference between the predecessor request Ra and the successor request Rb exceeds 250ms, the merge opportunity will be lost. In the environment of two-level disk scheduling, if the upstream scheduling performs too many request merging operations, the I/O requests arriving at the downstream scheduling may combine more sub-requests. However, due to the existence of A certain delay (assuming 10ms), so if the subsequent request Rb merges 25 I/O requests before reaching the downstream scheduling, then when Rb is sent to the downstream scheduling, Rb will lose the opportunity to merge with the predecessor request Ra. Moreover, since the delay overhead of virtualization needs to be considered when sending I/O requests from the virtual machine to the driver domain, when such "over-merging" occurs in the upstream scheduling, the I/O request merging ability of the downstream scheduling will drop sharply , thereby affecting disk I/O performance.

"Insufficient Merge":

In addition, if the amount of merging of application I/O requests is too large, the performance of I/O request merging will also be affected. For example, I/O requests R1, R2, ..., R50 arrive at the disk accordingly (the maximum time limit for I/O requests to reside is 250ms), the difference between two adjacent I/O requests is 10ms, and the 50 I/O requests All can be merged, then the first 25 I/O requests can be merged into I/O request Ra, but when R26 arrives at disk scheduling, I/O request Ra has been sent to the disk device. This means that I/O requests after R26 cannot continue to be merged with Ra. In the two-level disk scheduling process, if the upstream disk scheduling completes too few I/O requests to merge, then a large number of mergeable I/O requests will be sent to the downstream disk scheduling for further merge optimization. As mentioned above, the I/O request merging ability of downstream scheduling is limited, and the delay overhead of I/O requests sent from the virtual machine to the drive domain is not negligible. Then, the merging of I/O requests Optimization will suffer severely due to "under-merge" of upstream scheduling.

In order to clearly understand the present invention, relevant terms are explained below:

Driver Domain: The operating system that actually deploys the driver and is monitored by the virtualization system.

Front Driver: The virtualization driver deployed in the virtual machine is responsible for communicating with the driver domain to transmit the I/O request from the virtual machine side, and returning the execution result of the I/O request from the driver domain.

Back-end driver (Back Driver): The virtualization driver deployed in the drive domain, responsible for communicating with the virtual machine to transmit the execution result of the I/O request from the external device, and returning the I/O request from the virtual machine side.

Native Driver: Device intrinsic driver deployed in the driver domain.

Contents of the invention

The present invention provides a two-level disk scheduling method for a cloud computing environment, which reduces the uncoordinated problem of two-level disk scheduling existing in the existing two-level disk scheduling method, and reduces disk scheduling in the drive domain due to "excessive disk scheduling" in virtual machines. Merge" or "insufficient merge" has an adverse effect on the merge of I/O requests, thereby improving the overall storage performance of the cloud computing system.

A two-stage disk scheduling method oriented to a cloud computing environment provided by the present invention includes an initialization step, a monitoring step, a prediction step and a decision-making step:

(1) The initialization step includes the following sub-steps:

(1.1) Set parameter values:

Set the collection interval T, T is 0.5 seconds to 10 seconds;

Set the combination rate maximum limit value P _max and the combination rate average limit value P _ave , P _max is 0.4~0.6, and P _ave is 0~0.2;

Set the standard value of the merger termination rate σ _std , σ _std is 0.05~0.1;

Set the size b of the upstream scheduling arrival I/O request to 8 to 20 blocks, each block is 512 bytes;

Set the change step size u ₁ =4~11 of the first level combination limit value α;

Set the value range of α to [a ₁ , a ₁ +(n ₁ -1)×u ₁ ], where the lower limit a ₁ of the value range of α is 8 to 22, and the maximum number of steps n ₁ for α growth is symbol Indicates to find an integer upper limit for C;

Set the upper limit of the adjustment range α _max and the lower limit of the adjustment range α _{min of} α; α _max is 88~110, and α _min is 23~44;

Set the change step u ₂ of the size δ of the potential I/O request = 4 to 11;

Set the value range of δ to [a ₂ , a ₂ +(n ₂ -1)×u ₂ ], where the lower limit a ₂ of the value range of δ is 8 to 22, and the maximum number of steps n ₂ for δ change is

(1.2) Create a two-dimensional forecast table Two-dimensional index table and a one-dimensional index table 2D forecast table The first dimension of is The second dimension is the observation request size φ corresponding to α, and the entry is M _i,j , i=1~n ₁ , j=1~512; two-dimensional index table The first dimension of is The second dimension is Its value is the sequence number of φ corresponding to α, and the entry is Q _i,k , i=1~n ₁ , k=1~n ₂ ; one-dimensional index table The first dimension of is The entry is L _i , i=1~n ₁ ; the symbol Indicates to find the integer lower limit for C;

Clear all the entries in the above tables;

(1.3) set i=0, j=0;

(1.4) Judging whether a ₂ +(n ₂ -1)×u ₂ <512, if yes, the rotor step (1.5), otherwise the rotor step (1.15);

(1.5) Judging whether i<n ₁ , if yes, the rotor step (1.6), otherwise the rotor step (1.15);

(1.6) Judging whether k<n ₂ , if yes, the rotor step (1.7), otherwise the rotor step (1.14);

(1.7) Set α=a ₁ +i×u ₁ ;

(1.8) Set δ=a ₂ +k×u ₂ to calculate the potential merger rate p: p=1-b/δ;

(1.9) Calculate the observation combination rate m: m=1-(1-p)/(1-p ^α );

(1.10) Calculate the observation request size φ: Judging whether φ<512, if so, the rotor step (1.11), otherwise the rotor step (1.15);

(1.11) set j=φ, set M _i,j =δ, set Q _i,k =φ, set L _i =L _i +1;

(1.12) put k=k+1, rotor step (1.6);

(1.13) put i=i+1, rotor step (1.5);

(1.14) the initialization step ends, start step (2) and step (4);

(1.15) Abnormal exit;

(2) Monitoring step, including the following sub-steps:

(2.1) Setting the current moment t=0;

(2.2) Count the quantity A of upstream scheduling sending I/O requests; Statistical upstream scheduling sends the maximum value B ₁ and cumulative value B ₂ of the size of I/O requests; Statistical upstream scheduling merges the number of abort signals B ₃ ;

(2.3) Acquire time increment Δ, put t=t+Δ, judge whether t≥T, then rotor step (2.4), otherwise rotor step (2.2);

(2.4) Calculate the number of I/O requests sent by upstream scheduling per second S=A/T; calculate the average value B ₄ =B ₂ /A of the size of I/O requests sent by upstream scheduling; calculate the I/O requests sent by upstream scheduling The frequency B ₅ of the combined suspension signal = B ₃ /T;

(2.5) Set the maximum value of the observation request size φ ₁ =B ₁ , set the average value of the observation request size φ ₂ =B ₄ , and turn to the prediction step (3);

(2.6) judge whether to return abnormal mark, then rotor step (2.8), otherwise rotor step (2.7);

(2.7) Put the frequency ρ=B ₅ that upstream scheduling sends out the I/O request combination abort signal; Carry out sub-step (2.8);

(2.8) Set A=0, B ₁ =0, B ₂ =0, B ₃ =0, rotor step (2.1);

(3) Prediction step, including following sub-steps:

(3.1) Judging whether φ ₁ and φ ₂ are obtained successfully, if yes, the rotor step (3.2), otherwise the rotor step (3.15);

(3.2) Judging whether α<a ₁ , if yes, the rotor step (3.3), otherwise the rotor step (3.4);

(3.3) Set α=a ₁ ;

(3.4) Judging whether α>a ₁ +(n ₁ -1)×u ₁ , if yes, the rotor step (3.5), otherwise the rotor step (3.6);

(3.5) Set α=a ₁ +(n ₁ -1)×u ₁ ;

(3.6) put counter q=0,

(3.7) Set k=q, judge whether φ ₁ >Q _i,k and q<L _i , if yes, then rotor step (3.8), otherwise rotor step (3.9);

(3.8) put q=q+1, rotor step (3.7);

(3.9) Judging whether q≥L _i , if yes, the rotor step (3.15), otherwise the rotor step (3.10);

(3.10) Set j=Q _i,q , set the maximum predicted value of potential request size δ _max =M _i,j , set q=0;

(3.11) Set k=q, judge whether φ ₂ >Q _i,k and q<L _i , if yes, rotor step (3.12), otherwise rotor step (3.13);

(3.12) set q=q+1, rotor step (3.11);

(3.13) Judging whether q≥L _i , if yes, the rotor step (3.15), otherwise the rotor step (3.14);

(3.14) Set the average predicted value of the potential request size δ _ave = M _i,j ; if the prediction is successful, return to substep (2.6);

(3.15) put abnormal mark, return to substep (2.6);

(4) Decision-making step, including the following sub-steps:

(4.1) Judging whether δ _max and δ _ave change, if yes, then the rotor step (4.3), otherwise the rotor step (4.2);

(4.2) Wait for 2 ~ 10ms, the rotor step (4.1);

(4.3) Judging whether ρ and S change, if so, the rotor step (4.5), otherwise the rotor step (4.4);

(4.4) Wait for 2 ~ 10ms, the rotor step (4.3);

(4.5) Calculate the merger termination rate σ=ρ/S;

(4.6) Judging whether P _max >(δ _max -α)/δ _max and P _ave >(δ _ave -α)/δ _ave , if yes, the rotor step (4.7), otherwise the rotor step (4.10);

(4.7) Determine whether σ≤σ _std , if yes, the upstream scheduling is in the state of "over-merging", rotor step (4.8), otherwise rotor step (4.10);

(4.8) Determine whether α<α _min +u ₁ , if yes, go to the rotor step (4.1), otherwise go to the sub-step (4.9);

(4.9) Set α=α–u ₁ , rotor step (4.1);

(4.10) Determine whether P _max <(δ _max -α)/δ _max and σ>σ _std , if yes, the upstream scheduling is in the state of "insufficient merger", rotor step (4.12), otherwise rotor step (4.11);

(4.11) Judging whether P _ave <(δ _ave -α)/δ _ave and σ>σ _std , if yes, the upstream scheduling is in the state of "insufficient merger", rotor step (4.12), otherwise rotor step (4.1);

(4.12) Judging whether α>α _max -u ₁ , if yes, the rotor step (4.1), otherwise the rotor step (4.13);

(4.13) Set α=α+u ₁ , the rotor step (4.1).

In the initialization step, the setting of necessary parameters and the initialization of the two-dimensional prediction table M, the two-dimensional index table Q and the one-dimensional index table L are completed.

The monitoring step is to complete the real-time monitoring of I/O request merging related information.

The prediction step predicts the information reflecting the I/O request merging task amount according to the statistical information of the monitoring step, and transmits the prediction result to the monitoring step. In order to measure the extent to which disk scheduling performs I/O request merging, the present invention introduces an observed merging rate m, which is the ratio of the number of I/O request merging times to the number of arriving I/O requests. Since one I/O request merging operation can reduce the number of arriving I/O requests by one, the value range of the observed merging rate is [0, 1). For example, the combined I/O request size in formula (1) is θ:

θ=b/(1-m) (1)

In the formula, b represents the size of the upstream scheduled I/O request;

The potential merging rate p is the merging probability of I/O requests arriving at disk scheduling (the ratio of post-merging and pre-merging for arriving I/O requests), which can be obtained from formula (2):

p=1-b/δ (2)

In the formula, δ is the size of the potential I/O request, which is the maximum theoretical value of the average size of the merged I/O request;

According to the definition of the potential merger rate, the number E of I/O requests that can be merged can be obtained:

E=p×N (3)

Among them, N is the number of arriving I/O requests, so the potential merge rate p can indicate the amount of merged tasks for I/O requests.

Due to the limitation of the operating system on I/O request merging and the impact of disk scheduling on I/O request merging, the observed merging rate is less than or equal to the potential merging rate. In the case that the number of arriving I/O requests has been obtained, the potential merging rate p can be predicted according to the observed merging rate m, and the amount of merging tasks can also be obtained.

In order to effectively predict the potential merging rate and measure the amount of I/O request merging tasks, according to the law of I/O request merging, the request size θ after the limited I/O request merging:

θ＝b·(1+p+p ² +...+p ^α-1 ) (4)

Among them, α represents the first-level combined limit value;

From formulas (1) and (4), the corresponding observation combination rate m can be obtained:

m=1-(1-p)/(1-p ^α ) (5)

Because the I/O operation of the application is processed by the file system into an I/O request of a fixed size (determined by the physical page size of the operating system), and then sent to the disk for scheduling. Therefore, the initial size of the arriving I/O request in the disk scheduling (upstream scheduling) of the virtual machine is consistent (that is, b remains unchanged), so that the observed merge rate can be converted into the observed request size φ according to formula (1); Also convert the potential merge rate to the potential request size δ. In the implementation of the two-level disk scheduling optimization method, the relationship between φ and δ under different first-level combined limit values α is reflected in the form of a two-dimensional prediction table M. Figure 9 shows the corresponding relationship curve between potential request size and observed request size when α varies from 22 to 88.

The prediction process is a table look-up process. The index of the first dimension is the first-level combined limit value α, the index of the second dimension is φ, and the result of the detection is δ. The various parameters obtained in the prediction step are referenced by the monitoring step (ie δ _max and δ _ave ).

The decision-making step judges whether the upstream disk scheduling is in the "over-merging" state or "under-merging" state according to the parameters obtained in the monitoring step, and if so, dynamically adjusts the I/O request merging limit parameters of the upstream scheduling.

The parameters obtained by the decision-making step from the monitoring step include the maximum predicted value of potential request size δ _max , the average predicted value of potential request size δ _ave , the merge abort rate σ, the frequency of merge abort signals ρ and the number S of I/O requests. The ratio of merging abort signal frequency ρ to S is used as the effect feedback of the first-stage merging limit value α, which is called the merging abort rate σ. In addition, the decision-making step also requires the following parameters to participate in the decision-making, including the maximum limit value of the merger rate P _max , the average limit value of the merger rate P _ave , and the decision-making step sets the standard value of the merger termination rate σ _std and the first-level merger limit value α.

When the conditions described by formula (6) are met, the decision step considers the upstream scheduling to be in the state of "over-merging":

(P _max >(δ _max -α)/δ _max and P _ave >(δ _ave -α)/δ _ave ) and σ≤σ _std (6)

When the conditions described by formula (7) are met, the decision step considers the upstream scheduler to be in the "under-merge" state:

(P _max <(δ _max -α)/δ _max or P _ave <(δ _ave -α)/δ _ave ) and σ>σ _std (7)

When the decision-making step judges that the upstream scheduling is in the "over-merging" state, as long as the first-level merging limit value α is greater than α _min by one step size, α will be reduced by one step size u ₁ ; if the upstream scheduling is in the "under-merging" state, as long as the first-level If the first-level combined limit value α is less than α _max by one step, then α will be increased by one step u ₁ .

"Over-merging" and "under-merging" refer to the state that the upstream scheduler cannot coordinate with the downstream scheduler to complete the requested merge task. When the upstream scheduling is in the "over-merging" state, the delay of the I/O request in the upstream scheduling increases, thereby increasing the time interval between adjacent I/O requests arriving at the downstream scheduling, which will reduce the I/O request merging in the downstream scheduling In addition, when the upstream scheduling is in the "insufficient coalescing" state, the I/O requests are insufficiently merged in the upstream scheduling, thus causing more I/O requests to reach the downstream scheduling, and the I/O requests completed by the downstream scheduling within a limited time The amount of I/O request coalescing is limited, so more I/O requests that can be coalesced lose the opportunity to coalesce. Identifying and avoiding "over-coalescing" and "under-coalescing" states is an important optimization for request merging with two-level disk scheduling in cloud computing environments.

The present invention adopts the monitoring step to monitor in real time the information reflecting the request merging state, and provides the information to the decision-making step to judge whether the upstream scheduling is in the state of "excessive merging" or "insufficient merging". When it is judged that the upstream scheduling is "over-merging", the decision-making step reduces the first-level merging limit value α, thereby reducing the amount of merging of I/O requests in upstream scheduling, thereby reducing the time for adjacent I/O requests to reach downstream scheduling Time interval; when it is judged that the upstream scheduling is "insufficient in merging", the decision-making step increases the first-level merging limit value α to increase the amount of I/O request merging in the upstream scheduling, thereby reducing the number of merging I/O requests that are lost in the downstream scheduling. probability of opportunity. The monitoring step and the decision-making step cooperate with each other to form a feedback control, which reduces the uncoordinated problem of the two-level disk scheduling existing in the existing two-level disk scheduling method, and reduces the drive domain disk scheduling because of the "over-merging" or "merging" of the disk scheduling in the virtual machine. Insufficient" brings adverse effects on request merging, thereby improving the overall storage performance of the cloud computing system.

Description of drawings

Fig. 1 is the schematic diagram of drive separation model;

Figure 2 is a diagram illustrating the request for merging;

Fig. 3 is a request merging explanatory diagram of two-level disk scheduling;

Fig. 4 is a schematic flow chart of the present invention;

Figure 5 is a flowchart of the initialization steps.

Figure 6 is a flow chart of the monitoring steps.

Fig. 7 is a flow chart of the prediction steps.

Figure 8 is a flowchart of the decision-making steps.

Fig. 9 is a graph showing the relationship between the observed request size and the potential request size under the first-level combined limit value α.

Detailed ways

Further description will be given below in conjunction with drawings and embodiments.

As shown in Fig. 4, the embodiment of the present invention includes an initialization step, a monitoring step, a prediction step and a decision-making step.

The initialization steps, monitoring steps, prediction steps, and decision steps are shown in Figure 5, Figure 6, Figure 7, and Figure 8, respectively.

In this embodiment, the substep (1.1) and substep (1.2) of the initialization step are respectively:

(1.1) Set parameter values:

Set the collection interval T, T is 1 second;

Set the maximum limit value of the merger rate P _max and the average limit value of the merger rate P _ave , P _max is 0.5, and _Pave is 0.1;

Set the standard value of the merger suspension rate σ _std , σ _std is 0.05;

Set the size b of the upstream scheduled I/O request to 8 blocks, each block is 512 bytes;

Set the change step size u ₁ =11 of the first level combined limit value α;

Set the value range of α to [11, 132], where the lower limit a ₁ of the value range of α is 11, and the maximum number of steps n ₁ for α growth is 12;

Set the upper limit of the adjustment range α _max and the lower limit of the adjustment range α _min of α; α _max is 110, and α _min is 23;

Set the change step size u ₂ =11 of the size δ of the potential I/O request;

Set the value range of δ to [11,506], wherein the lower limit _a2 of the value range of δ is 11, and the maximum number of steps _n2 of δ change is 46;

(1.2) Create a two-dimensional forecast table Two-dimensional index table and a one-dimensional index table 2D forecast table The first dimension of is The second dimension is the observation request size φ corresponding to α, the entry is M _i,j , i=1~12, j=1~512; two-dimensional index table The first dimension is (α-a ₁ )/u ₁ , the second dimension is the serial number of α corresponding to φ, and the entry is Q _i,k , i=1~12, k=1~46; one-dimensional index table The first dimension of is The entry is L _i , i=1~12;

Clear all entries in the above tables to zero.

Claims (1)

1. a two-stage disk-scheduling method for facing cloud computing environment, comprises initialization step, monitoring step, prediction steps and steps in decision-making:

(1) initialization step comprises following sub-step:

(1.1) parameters value:

Arrange acquisition interval T, T is 0.5 second ~ 10 seconds;

Merger ratio maximum limit definite value P is set _maxlimit value P average with merger ratio _ave, P _maxbe 0.4 ~ 0.6, P _avebe 0 ~ 0.2;

Arrange and merge termination rate standard value σ _std, σ _stdbe 0.05 ~ 0.1;

The size b arranging upstream scheduling arrival I/O request is 8 pieces ~ 20 pieces, every block 512 byte;

The change step u that the first order merges limits value α is set ₁=4 ~ 11;

The span of α is set for [a ₁, a ₁+ (n ₁-1) × u ₁], wherein, the span lower limit a of α ₁be the maximum step number n that 8 ~ 22, α increases ₁for symbol represent and the integer upper limit is asked to C;

The setting range upper limit α of α is set _maxwith setting range lower limit α _min; α _maxbe 88 ~ 110, α _minbe 23 ~ 44;

The change step u of the size δ that potential I/O asks is set ₂=4 ~ 11;

The span of δ is set for [a ₂, a ₂+ (n ₂-1) × u ₂], wherein, the lower limit a of the span of δ ₂be the maximum step number n of 8 ~ 22, δ change ₂for

(1.2) two-dimensional prediction table is created 2-d index table with one dimension concordance list two-dimensional prediction table first dimension be second dimension is the observation request size φ that α is corresponding, and list item is M _{i, j}, i=1 ~ n ₁, j=1 ~ 512; 2-d index table first dimension be second dimension is its value is the sequence number of the corresponding φ of α, and list item is Q _{i, k}, i=1 ~ n ₁, k=1 ~ n ₂; One dimension concordance list first dimension be list item is L _i, i=1 ~ n ₁; Symbol represent and integer lower limit is asked to C;

The list item of above-mentioned each table is all reset;

(1.3) i=0 is put, j=0;

(1.4) a is judged whether ₂+ (n ₂-1) × u ₂<512 is then rotor step (1.5), otherwise rotor step (1.15);

(1.5) i<n is judged whether ₁, be then rotor step (1.6), otherwise rotor step (1.15);

(1.6) k<n is judged whether ₂, be then rotor step (1.7), otherwise rotor step (1.14);

(1.7) α=a is put ₁+ i × u ₁;

(1.8) δ=a is put ₂+ k × u ₂, calculate potential merger ratio p:p=1-b/ δ;

(1.9) calculating observation merger ratio m:m=1-(1-p)/(1-p ^α);

(1.10) calculating observation request size φ: judging whether φ <512, is then rotor step (1.11), otherwise rotor step (1.15);

(1.11) put j=φ, put M _{i, j}=δ, puts Q _{i, k}=φ, puts L _j=L _i+ 1;

(1.12) k=k+1 is put, rotor step (1.6);

(1.13) i=i+1 is put, rotor step (1.5);

(1.14) initialization step terminates, setting up procedure (2) and step (4);

(1.15) extremely exit;

(2) monitoring step, comprises following sub-step:

(2.1) current time t=0 is set;

(2.2) the quantity A that upstream scheduling sends I/O request is added up; The scheduling of statistics upstream sends the maximal value B that I/O asks size ₁with accumulated value B ₂; The merging abort signal quantity B of statistics upstream scheduling ₃;

(2.3) acquisition time increment Delta, puts t=t+ Δ, judges whether t >=T, is then rotor step (2.4), otherwise rotor step (2.2);

(2.4) the quantity S=A/T that upstream per second scheduling sends I/O request is calculated; Calculate upstream scheduling and send the mean value B that I/O asks size ₄=B ₂/ A; Calculate upstream scheduling and send the frequency B merging abort signal in I/O request ₅=B ₃/ T;

(2.5) the maximal value φ of observation request size is put ₁=B ₁, put the mean value φ of observation request size ₂=B ₄, turn prediction steps (3);

(2.6) judging whether to return abnormal marking, is then rotor step (2.8), otherwise rotor step (2.7);

(2.7) frequency ρ=B that trip scheduling sends I/O request merging abort signal is set up ₅; Carry out sub-step (2.8);

(2.8) A=0 is put, B ₁=0, B ₂=0, B ₃=0, rotor step (2.1);

(3) prediction steps, comprises following sub-step:

(3.1) judge whether successfully to obtain φ ₁and φ ₂, be then rotor step (3.2), otherwise rotor step (3.15);

(3.2) α <a is judged whether ₁, be then rotor step (3.3), otherwise rotor step (3.4);

(3.3) α=a is put ₁;

(3.4) α >a is judged whether ₁+ (n ₁-1) × u ₁, be then rotor step (3.5), otherwise rotor step (3.6);

(3.5) α=a is put ₁+ (n ₁-1) × u ₁;

(3.6) counter q=0 is put,

(3.7) put k=q, judge whether φ ₁>Q _{i, k}and q<L _i, be then rotor step (3.8), otherwise rotor step (3.9);

(3.8) q=q+1 is put, rotor step (3.7);

(3.9) q>=L is judged whether _i, be then rotor step (3.15), otherwise rotor step (3.10);

(3.10) j=Q is put _{i, q}, put the maximum predicted value δ of potential request size _max=M _{i, j}, put q=0;

(3.11) put k=q, judge whether φ ₂>Q _{i, k}and q<L _i, be then rotor step (3.12), otherwise rotor step (3.13);

(3.12) q=q+1 is set, rotor step (3.11);

(3.13) q>=L is judged whether _i, be then rotor step (3.15), otherwise rotor step (3.14);

(3.14) the average magnitude predicted value δ of potential request size is put _ave=M _{i, j}; Predict successfully, return sub-step (2.6);

(3.15) put abnormal marking, return sub-step (2.6);

(4) steps in decision-making, comprises following sub-step:

(4.1) δ is judged _maxand δ _avewhether change, be then rotor step (4.3), otherwise rotor step (4.2);

(4.2) 2 ~ 10ms is waited for, rotor step (4.1);

(4.3) judging whether ρ and S changes, is then rotor step (4.5), otherwise rotor step (4.4);

(4.4) 2 ~ 10ms is waited for, rotor step (4.3);

(4.5) merging termination rate σ=ρ/S is calculated;

(4.6) P is judged whether _max> (δ _max-α)/δ _maxand P _ave> (δ _ave-α)/δ _ave, be then rotor step (4.7), otherwise rotor step (4.10);

(4.7) σ≤σ is judged whether _std, be, upstream scheduling is in " excessively merging " state, rotor step (4.8), otherwise rotor step (4.10);

(4.8) α < α is judged whether _min+ u ₁, be then rotor step (4.1), otherwise carry out sub-step (4.9);

(4.9) α=α – u is put ₁, rotor step (4.1);

(4.10) P is judged whether _max< (δ _max-α)/δ _maxand σ > σ _std, be, upstream scheduling is in " merging not enough " state, rotor step (4.12), otherwise rotor step (4.11);

(4.11) P is judged whether _ave< (δ _ave-α)/δ _aveand σ > σ _std, be, upstream scheduling is in " merging not enough " state, rotor step (4.12), otherwise rotor step (4.1);

(4.12) α > α is judged whether _max-u ₁, be then rotor step (4.1), otherwise rotor step (4.13);

(4.13) α=α+u is put ₁, rotor step (4.1).