CN102739785A - Method for scheduling cloud computing tasks based on network bandwidth estimation - Google Patents

Abstract

The invention discloses a method for scheduling cloud computing tasks based on network bandwidth estimation. The platform of the method is divided into three stages which are a data center, shelves and compute nodes, and is uniformly managed by a management node, and the compute nodes are used for executing the tasks and acquiring available bandwidth information; when executing tasks, the compute nodes read input data from a local disk or by a network, and when the tasks are finished, the compute nodes request for new tasks from the management node; the management node is used for managing the available bandwidth information and scheduling the tasks, when managing the available bandwidth information, the management node collects bandwidth data of the compute nodes, specifies two compute nodes in different shelves to execute available bandwidth estimation programs, and counts the available bandwidth information; and when the compute nodes request for the tasks, the management node determines a task scheduling strategy by combining network bandwidths in the shelves, the available bandwidth between the shelves and load of the data center.

Description

The cloud computing method for scheduling task of bandwidth estimation Network Based

Technical field

The present invention relates to computer network field, particularly task scheduling technology is specifically related to a kind of cloud computing method for scheduling task of bandwidth perception Network Based.

Background technology

Along with the extensive use of mass data processing (like Web log analysis, OLAP, OLTP), rapid towards cloud computing platform (like MapReduce, Hadoop, the Dryad etc.) development of mass data processing.

These cloud computing platforms shared not have (Shared Nothing) Business Information and IT Solution Mgmt Dep is deployed on data center.Aspect storage, the some blocks of files with sizes such as data file are divided into are stored in the local hard drive of computing node, and utilize distributed file system to manage.In order to guarantee high reliability, each blocks of files has a plurality of copies.Aspect calculating; Data processing operation (Job) is divided into some independently subtasks (Task), and each task is from the computing node local hard drive or pass through read input file piece, and handles; After accomplished all subtasks, data processing operation could be accomplished.Utilize the distributed storage of data and the parallelization of task to carry out, the TB level can read and write and handle to cloud computing platform efficiently, even the mass data of PB level.

But the network bandwidth is data center's resource in short supply, and has a strong impact on the cloud computing platform performance.At first, in order to keep high performance-price ratio, the computing node of most of data centers does not adopt high-speed network appliance (like ten thousand mbit ethernets) interconnected; Secondly, the network topology (can be divided into three grades of data centers, frame, computing node) of hierarchy type is adopted at the partial data center, has bandwidth bottleneck during computing node communication in the different frames; Once more, the part calculation task need pass through the read data, and the Network Transmission expense influences the task deadline.Action needs such as the read-write of large-scale data at last,, blocks of files copy creating and migration take a large amount of network bandwidths.

Summary of the invention

The present invention is directed to the deficiency and the defective of prior art; A kind of cloud computing method for scheduling task of bandwidth estimation Network Based has been proposed; This method is through the operating load and the network availability bandwidth at perception data center; Dynamically adjust the scheduling strategy of scheduler, with the time of implementation of minimizing operation, thus the execution performance of raising cloud computing system.

The technical scheme that the present invention adopts: a kind of cloud computing method for scheduling task of bandwidth estimation Network Based may further comprise the steps:

A. each computing node at first utilizes the networking command (like ifconfig, sar etc.) of operating system whenever to read the flow of network adapter at a distance from 3 seconds; Calculate the available bandwidth of this computing node again, the bandwidth information with this computing node sends to management node at last.

B. the bandwidth capture program of management node is collected the available bandwidth information of each computing node, and in internal memory, preserves the bandwidth information in nearest 1 minute of each computing node, and calculates the average B of the available bandwidth of nearest 1 minute computing node j _jAverage with all computing node available bandwidths

Clear up expired (before the 5 minutes) data in the internal memory at last.

C. management node is selected a highest computing node of average available bandwidth from each frame, makes these computing nodes move estimation of available bandwidth program (pathchirp) between any two.Available bandwidth B between frame m and frame n ^{M, n}Available above-mentioned estimated result replaces.The average of available bandwidth between frame

D. after computing node j finishes the work, please looking for novelty to management node of task.

E. management node writes down current time T after the request of receiving computing node, and adds up the quantity of the request of receiving in nearest a minute, calculates the request of each second at last and counts λ, be i.e. request frequency.

F. management node selects suitable task scheduling to computing node j according to local task timer and frame task timer.

1) upgrades local task timer and frame task timer, utilize towards the delay scheduling model computing relay of hierachical network topology and wait for window

and

2) from task queue, search the task of input data on this computing node,, then be dispatched to computing node j,, change 7) local task timer and the zero clearing of frame task timer if find;

3) more local task timer postpones to wait for window if the local task timer time is not then dispatched any task to computing node j less than

with local, changes 7);

4) from task queue, search the task of input data,, then be dispatched to computing node j,, change 7) the zero clearing of frame task timer if find in computing node j place frame;

5) relatively the frame task timer postpones to wait for window

if the frame task timer time is not then dispatched any task to computing node j less than

with local, changes 7);

6) from task queue, choose a task scheduling wantonly to computing node j;

7) finish

Delay scheduling model towards hierachical network topology is following:

Make the blocks of files size of input file be F; Can find the probability of the task of input data on computing node j in the task queue is p ^d, can find input data is p at the probability of the task of the frame at computing node j place ^r, p ^l=p ^d+ p ^rThe input data are in the estimated time of frame internal transmission

The estimated time that the input data are transmitted between frame

1) when $T_{\mathrm{Core}}\≤\frac{1}{p^l\mathrm{\λ}}+\frac{p^r}{p^l}{T}_{\mathrm{Rack}}$ The time, $T_d^w=0$ And $T_r^w=0.$

2) when $T_{\mathrm{Core}}>\frac{1}{p^l\mathrm{\λ}}+\frac{p^r}{p^l}{T}_{\mathrm{Rack}}$ And $T_{\mathrm{Rack}}>\frac{1}{p^d\mathrm{\λ}}$ The time, $T_d^w=0,$

$T_r^w=\left\{\begin{array}{cc}-\frac{\ln Q^{\&prime;}}{p^l\mathrm{\&lambda;}}& Q^{\&prime;}>1\\ 0& Q^{\&prime;}<1\end{array}\right.$

Wherein, $Q^{\&prime;}=\frac{\mathrm{\&alpha;}(\frac{1}{p^l\mathrm{\&lambda;}}+\frac{p^r}{p^l}{T}_{\mathrm{Rack}})}{(1-p^l)(T_{\mathrm{Core}}-\frac{1}{p^l\mathrm{\&lambda;}}-\frac{p^r}{p^l}{T}_{\mathrm{Rack}})},$ 0<α<1.

3) when $T_{\mathrm{Core}}>\frac{1}{p^l\mathrm{\&lambda;}}+\frac{p^r}{p^l}{T}_{\mathrm{Rack}}$ And $T_{\mathrm{Rack}}>\frac{1}{p^d\mathrm{\&lambda;}}$ The time, calculate order for ease $T_r^w=T_d^w.$

$\left((1-p^l)(T_{\mathrm{core}}-\frac{1}{p^l\mathrm{\&lambda;}}-\frac{p^r}{p^l}{T}_{\mathrm{rack}})\right)e^{-(p^l+p^d)\mathrm{\&lambda;}{T}_d^w}$

$+{(\frac{1}{p^l\mathrm{\&lambda;}}+\frac{p^r}{p^l}{T}_{\mathrm{rack}}-\frac{1}{p^d\mathrm{\&lambda;}})e}^{-p^d\mathrm{\&lambda;}{T}_d^w}=\frac{\mathrm{\&alpha;}}{p^d\mathrm{\&lambda;}}$

Wherein, α (0 < α < 1) is for waiting for the tolerance factor, and the time of the more little wait of α is long more, owing to above-mentioned formula is monotonic function and exists rational to separate, so can adopt Newton iteration method to find the solution.When solving The time, order $T_d^w=0.$

G. after computing node j receives the scheduling decision of management node, do following reaction:

1) if management node does not arrive computing node j with any task scheduling, only if there is task to accomplish the task that computing node j no longer please look for novelty in 3 seconds from now on.

2) if management node is dispatched to computing node j with task i, then computing node j is according to the position of task i input deposit data, from local hard drive, read the input data with other computing nodes of frame or the computing node of other frames.

Network bandwidth estimation of the present invention aspect, each computing node reads the load of network adapter in real time from operating system, and reports to management node; Management node is regularly selected the highest computing node of the network bandwidth from each frame, and arranges operation estimation of available bandwidth program between these computing nodes, to gather the available bandwidth between frame; The task scheduling aspect; Management node at computing node to its request during task; Read the frequency of network bandwidth information and idle computing node request task earlier, again the estimation tasks deadline, select a suitable task scheduling at last to this computing node; Thereby when task scheduling, consider network bandwidth factor, reach the purpose that improves the cloud computing system overall performance.

Beneficial effect of the present invention:

(1) the real-time network bandwidth information of each computing node of image data center, but and the available bandwidth between the approximate evaluation frame, help the network condition situation that the keeper grasps data center intuitively.

(2) network bandwidth monitoring framework can need not to change any hardware device based on the existing device build of data center, and can monitor the available bandwidth of each computing node and the available bandwidth between frame.

(3) dispatching algorithm is considered the load and the network bandwidth factor of data center simultaneously towards the network topology of layering, avoids causing systematic function to reduce because of waits for too long or network latency are long.

Description of drawings

Fig. 1 is the data center's sketch map that the present invention is based on the hierarchical network topology;

Fig. 2 is the flow chart of network bandwidth method of estimation of the present invention;

Fig. 3 is the cloud computing task scheduling algorithm flow chart of bandwidth estimation Network Based of the present invention.

Embodiment:

Below in conjunction with accompanying drawing and embodiment the present invention is remake further detailed explanation.

The present invention moves under the cloud computing environment with hierarchical network topological structure, and is as shown in Figure 1.This cloud computing environment is divided into three grades of node, frame and data centers: deployment computing node and frame switch in the frame, and computing node links to each other with the frame switch; Several frame composition data centers link to each other through core switch between the frame.This cloud computing environment also comprises a management node, and management node links to each other with computing node with the frame switch through core switch.Computing node is responsible for executing the task, the monitor network bandwidth, and whenever report band data to computing node at a distance from 3 seconds.Management node is responsible for some tasks are resolved in the operation that the user submits to, and these tasks is submitted to computing node carry out.After receiving the node available bandwidth information of computing node collection; Management node is written into Bandwidth Broker and database with the available bandwidth information of each node; Bandwidth Broker calculates the average available bandwidth in nearest 1 minute of each computing node respectively, and the band data of expired (before 5 minutes) in the cleaning internal memory.Every at a distance from 1 minute, management node is chosen the maximum computing node of available bandwidth from each frame, and makes these nodes carry out estimation of available bandwidth program (like pathchirp) between any two, to estimate the available bandwidth between the frame.After a cloud computing operation arrived, management node was divided into plurality of sub task (being called for short " task ") with this operation, and deposits task pool in.After the task actuator of a computing node is finished the work; The task that this computing node please be looked for novelty to management node, the task dispatcher of management node selects a suitable task scheduling to this computing node (or temporarily not scheduler task) from task pool according to the network bandwidth and data center's load.The task actuator of computing node is according to the requirement of task to the input data, reading of data from the storage of the storage of this locality or other computing nodes.Need read blocks of files 1 such as hypothesis task 1, task 2 need read blocks of files 2.If computing node 1 need execute the task 1, then reading of data from local disk; If computing node 1 need execute the task 1, then through Network Transmission, from other computing nodes (like computing node nm) reading of data.

The concrete steps of network bandwidth monitoring are as shown in Figure 2, two frames arranged, four computing nodes (two computing nodes of each frame), a management node in this cloud computing environment.

(1) flow of network adapter is read in the order of each computing node call operation system (like orders such as ifconfi g, sar), and whenever at a distance from 3 seconds available bandwidth information is sent to management node.

(2) safeguard a Hash table HashTable < Key, Value>in the internal memory of management node, Key is the computing node name, and Value is the chained list of storage available bandwidth value.After management node is received the available bandwidth information of a computing node; From Hash table, take out the corresponding chained list of this node; Current time and available bandwidth value are joined in this chained list, simultaneously expired available bandwidth value is write database, and in internal memory, delete these values.

(3) management node is added up the average available bandwidth value in nearest a minute for each computing node, and from each frame, chooses a node formation frame collection that average available bandwidth is the highest.

(4) management node makes the computing node in the frame collection carry out estimation of available bandwidth program (pathchirp) between any two, the result of management node collection monitoring, and the mean value of bandwidth between the computer rack.The task scheduling pattern is as shown in Figure 3,

(1) when computing node 1 was idle, this node was to management node request task.

(2) management node is according to the scheduling result assign task

(3) computing node 1 is carried out the task of assigning

(4) after computing node 1 is finished the work, please looking for novelty of task

(5) computing node 1 is to the data of computing node 2 request required by task

(6) computing node 1 is received after the data of required by task through network, begins to execute the task

(7) computing node 1 is finished the work

Management node supposes to have data center that 10 frames are arranged according to the scheduling result assign task, 10 computing nodes of each frame; Totally 100 computing nodes have 5 tasks in the task pool, and each task need read a blocks of files; Each blocks of files has 3 copies; 5 computing node request tasks are arranged each second, and the average of data transmission period is 1 second in the frame, and striding the frame data transmission period is 10 seconds.Then, the Probability p of required data block is arranged on its local hard drive for a computing node ^d=1-(1-3/100) ⁵≈ 0.14, and the Probability p of required data block is arranged on its place other nodes of frame ^r=1-((1-3/100) ⁵) ¹⁰-p ^d≈ 0.64, makes p ^l=p ^d+ p ^r≈ 0.78.

When computing node S was idle, scheduler judged whether the task of data on S earlier, if having, then with this task scheduling to S, and with local task timer zero clearing, otherwise do following calculating.

1) calculates Because $T_{\mathrm{Core}}>\frac{1}{p^l\mathrm{\&lambda;}}+\frac{p^r}{p^l}{T}_{\mathrm{Rack}}=\frac{1}{0.78\&times;5}+\frac{0.64}{0.78}\&times;1=1.07,$ $T_{\mathrm{Rack}}\&le;\frac{1}{p^d\mathrm{\&lambda;}}=1.42,$ So $T_d^w=0$

2) calculate and make α=0.01, then

$Q^{\&prime;}=\frac{\mathrm{\&alpha;}(\frac{1}{p^l\mathrm{\&lambda;}}+\frac{p^r}{p^l}{T}_{\mathrm{Rack}})}{(1-p^l)(T_{\mathrm{Core}}-\frac{1}{p^l\mathrm{\&lambda;}}-\frac{p^r}{p^l}{T}_{\mathrm{Rack}})}=\frac{0.01\&times;1.07}{0.22\&times;(2-1.07)}=0.0054,$ Can get

$T_r^w=-\frac{\ln Q^{\&prime;}}{p^l\mathrm{\&lambda;}}\&ap;1.34$

Because the local wait window

that postpones contains the computing node of desired data so scheduler is no longer waited for local hard drive.Scheduler is sought the task of data other nodes of frame at S place from task pool, if having, then with this task scheduling to S, otherwise upgrade the frame task timer.If the frame task timer postpones to wait for the then optional task scheduling of window

to S greater than frame, otherwise does not do any scheduling.

Should be pointed out that for those skilled in the art, under the prerequisite that does not break away from the principle of the invention, can also make some improvement and retouching, these improvement and retouching also should be regarded as protection scope of the present invention.The all available prior art of each part not clear and definite in the present embodiment realizes.

Claims (5)

1. the cloud computing method for scheduling task of a bandwidth estimation Network Based, it is characterized in that: this method may further comprise the steps:

Step 1: computing node dynamic monitoring meshed network bandwidth, and regularly to management node report bandwidth information; The average bandwidth of management node computational analysis node, and for suitable computing node of each frame election to measure the bandwidth between the frame; The computing node of being elected is carried out the estimation of available bandwidth program, and bandwidth estimation result is reported to management node; Management node is analyzed the average bandwidth between frame;

Step 2: after computing node is finished the work, please looking for novelty to management node of task; Management node upgrades local task timer and frame task timer, and according to the arrival intensity of idle computing node, the operating load at calculated data center;

Step 3: the task in the management node scan task pond, priority scheduling data are in the task of current idle computing node local hard drive, if there is not such task, then computing relay is waited for window, and makes scheduling decision according to window size;

Step 4: management node according to scheduling result with local task timer or the zero clearing of frame task timer.

2. the cloud computing method for scheduling task of bandwidth estimation Network Based according to claim 1; It is characterized in that: the available bandwidth described in the said step 1 is the data that system gathers in real time; Computing node whenever utilizes the operating position of the networking command collection network adapter of operating system at a distance from 3 seconds, and calculate available bandwidth; After management node is received the bandwidth information from each computing node; Earlier deposit it in internal memory and database; Ask bandwidth average in nearest one minute of each computing node again, add up bandwidth average in nearest one minute of all computing node bandwidth then, clear up the stale data in the internal memory at last; Management node was whenever selected the maximum computing node of bandwidth at a distance from 1 minute for each frame; Selecteed computing node utilizes the estimation of available bandwidth program to estimate available bandwidth between any two, and the result is sent to management node; Management node is added up bandwidth information between frame, and obtains the average of available bandwidth between frame.

3. the cloud computing method for scheduling task of bandwidth estimation Network Based according to claim 1; It is characterized in that: local task timer described in the said step 2 and frame task timer are two numerical value, write down the current time and the timer zero clearing last time time interval constantly respectively; The operating load of data center is the frequency of computing node request task; The high more expression operating load of the frequency of computing node request task is low more, and vice versa; The frequency of computing node request task is tried to achieve through the method for statistics; Whenever a computing node free time, the task that this computing node please be looked for novelty to management node, management node writes down current time point, and adds up the number of the computing node of 1 fen interior request task of clock time, thereby obtains the frequency of computing node request task.

4. the cloud computing method for scheduling task of bandwidth estimation Network Based according to claim 1; It is characterized in that: the delay wait window described in the said step 3 is tried to achieve by the delay scheduling model towards hierachical network topology, postpones to wait for that window is divided into local delay wait window and frame postpones to wait for two kinds of windows; Scheduler elder generation is according to the operating load and the network bandwidth information of data center; Faying face is to the delay scheduling model of hierachical network topology; Calculate local delay and wait for that window and frame postpone to wait for the value of window; For the computing node of current request task, if the required data of task K in the task pool then are dispatched to this computing node with task K on this node; Otherwise judge whether local timer postpones to wait for window less than this locality, if less than, any scheduling then do not done; Otherwise the data whether required by task is arranged in the inspection task pool if can find such task, then arrive this computing node with this task scheduling on the frame at this computing node place; If can not find suitable task, judge then whether this frame timer postpones to wait for window less than frame, if less than, then do not do any scheduling; Otherwise select a task at random to this computing node.

5. the cloud computing method for scheduling task of bandwidth estimation Network Based according to claim 1; It is characterized in that: the reason node described in the said step 4 is meant as if the data of having dispatched a task and this required by task local task timer or the zero clearing of frame task timer at the computing node local hard drive, then with local task timer and the zero clearing of frame task delay timer according to scheduling result; Refer to if dispatched the data of a task and this required by task other nodes of frame, then with the zero clearing of frame task delay timer at the computing node place; If the data of the task of scheduling are in other frames, then not zero clearing; If do not work as dispatcher then not zero clearing.

Images