CN104902001A - Method for load balancing of Web requests based on operating system virtualization - Google Patents

Abstract

The invention discloses a method for load balancing of Web requests based on operating system virtualization. The method comprises the following steps of firstly, carrying out collection and normalization of server resource information; secondly, collecting dynamic information of a server and service instances; thirdly, computing final weights of service instance copies of all services according to the server resource information and the periodically-collected dynamic information; and lastly, distributing the requests through weighted round robin. According to the method, the phenomenon that the different service instance copies on one server simultaneously receive a large amount of Web requests due to the lower server load to cause the server overload can be effectively avoided, the load balancing between the service instance copies and the server is implemented, the concurrence capacity of a Web cluster system is improved, and the average response time of the requests is reduced.

Description

Load-balancing method is asked based on the virtualized Web of operating system

Technical field

The invention belongs to Web cluster load balance technical field, be specifically related to a kind of based on operating system virtualized Web request load-balancing method.

Background technology

Along with the fast development of the internet, applications such as social networks, ecommerce, the webserver is faced with the problem of following two aspects: (1) access number exponentially doubly increases; (2) transaction is more complicated.In order to tackle above problem, for user provides a high performance web services environment, distributed Web server system (Web cluster) is arisen at the historic moment, and in order to replace that there is high performance separate unit Web server.In Web cluster, every class services package is containing multiple Service Instance copy, and multiple Service Instance copies of same service are carried by multiple different server.In order to user's request reasonably, is balancedly assigned to different Service Instance copies, guarantee the load balance between each server and Service Instance copy, improve the concurrent capability of Web group system and the utilance of system resource, efficient load balancing is the key point of problem.

Tradition Web cluster carries Service Instance copy in units of physical machine or virtual machine, and according to load balancing, user's request is reasonably distributed to the physical machine or virtual machine that carry different Service Instance copy.But, along with the development of cloud computing technology, occur based on the base unit of operating system virtualized Docker container as a kind of carrying Service Instance newly.Because Docker employs the Intel Virtualization Technology based on operating system, therefore, compared with traditional virtual machine based on Full-virtualization, Docker can realize runtime system resource elasticity distribution function (namely when in Docker container without tasks carrying time, Docker container does not take cpu resource; When having a tasks carrying, Docker container takies required cpu resource according to its CPU weight size), improve the utilance of server resource.These new features of Docker, make the load balancing on traditional Web cluster no longer applicable, therefore, need to build and a kind ofly rational, balanced ask load-balancing method based on the virtualized Web of operating system.

Summary of the invention

The present invention is directed to the deficiencies in the prior art, provide a kind of based on operating system virtualized Web request load-balancing method.

The concrete steps of the inventive method are:

Step (1) sets up server resource state information list.In Web cluster, server set S is expressed as:

S＝{s ₁,s ₂,s ₃,...,s _j,...,s _n}

Wherein s _j(1≤j≤n) represents a certain server in Web cluster, and n represents the server sum that Web cluster comprises.Server s _jtotal resource P _jbe expressed as:

P _j＝(P _{j_cpu},P _{j_memory},P _{j_io},P _{j_network})

Wherein P _{j_cpu}represent server s _jthe computing capability of CPU, P _{j_memory}represent server s _jinternal memory disposal ability, P _{j_io}represent server s _jhard disk I/O ability, P _{j_network}represent server s _jnetwork throughput.In order to the impact that the otherness eliminating server isomerism and variety classes resource is brought, adopt Max-Min method to server s _jtotal resource P _jvalue is normalized, server s _jtotal resource normalized value PO _jbe expressed as:

PO _j＝(PO _{j_cpu},PO _{j_memory},PO _{j_io},PO _{j_network})

Step (2) sets up the list of Service Instance resource state information.If the set of service F that Web cluster externally provides is expressed as:

F＝{f ₁,f ₂,f ₃,...,f _i,...f _m}

Wherein f _i(1≤i≤m) represents that i-th kind of service that cluster externally provides, m represent the type service sum that cluster provides.Service f _icomprise multiple different Service Instance copy set F _ibe expressed as:

F _i＝{f _i1,f _i2,f _i3,…,f _ik,…,f _il}

Wherein f _ik(1≤k≤l) represents service f _iexample copy, l represent service f _iexample copy number.Service Instance f _ikresource state information ST _ikbe expressed as:

${\mathrm{ST}}_{\mathrm{ik}}=(L_{\mathrm{ik}}^j,\mathrm{cpu}\_{\mathrm{share}}_{\mathrm{ik}},{\mathrm{memory}}_{\mathrm{ik}\_\mathrm{total}})$

Wherein represent service f _iexample copy f _ikthe main frame s at place _j, cpu_share _ikrepresent service f _iexample copy f _ikcPU weights, service f _iexample copy f _ikmaximum can committed memory.

Step (3) is every T cycle time, and load-balanced server can the load information of Servers-all in this cycle time T in periodic collection Web cluster.Basic load information comprises:

Server s _jcpu busy percentage:

CPU _j=cpu busy time/(cpu busy time+cpu idle time)

Server s _jhard disk I/O load:

IO _j=hard disk IO the rush hour/(hard disk IO rush hour the+hard disk IO free time)

Server s _joffered load:

Network _j=(flowing into output flow in flow+cycle T in cycle T)/(cycle T * P _{j_network})

Server s _jupper all example copy CPU weights number sums:

${\mathrm{\Σ}}_{\∀f_{\mathrm{ik}}\∈F_i,L_{\mathrm{ik}}^j\∈s_j}\mathrm{cpu}\_{\mathrm{share}}_{\mathrm{ik}}^j$

Server s _jupper service f _iexample copy f _ikmemory usage:

${\mathrm{Memory}}_{\mathrm{ik}}^j={\mathrm{memory}}_{\mathrm{ik}\_\mathrm{used}}/{\mathrm{memory}}_{\mathrm{ik}\_\mathrm{total}}$

Wherein memory _{ik_used}represent service f _iexample copy f _ikthe memory value used;

Step (4) load-balanced server is according to the server load information of collecting and existing service device and Service Instance state information, and calculating part is deployed in server s _jon Service Instance f _ikfinal weights and upgrade the value information of respective service example in weights list.

$\begin{array}{c}{\mathrm{Weight}}_{\mathrm{ik}}^j={\mathrm{\α}}_{\mathrm{ik}}*{\mathrm{PO}}_{j\_\mathrm{cpu}}*(1-{\mathrm{CPU}}_j)*\frac{{\mathrm{cpu}\_\mathrm{share}}_{\mathrm{ik}}^j}{{\mathrm{\Σ}}_{{\∀f}_{\mathrm{ik}}\∈F_i,L_{\mathrm{ik}}^j\∈s_j}{\mathrm{cpu}\_\mathrm{share}}_{\mathrm{ik}}^j}+\\ {\mathrm{\β}}_{\mathrm{ik}}*{\mathrm{PO}}_{j\_\mathrm{memory}}*(1-{\mathrm{Memory}}_{\mathrm{ik}}^j)+{\mathrm{\γ}}_{\mathrm{ik}}*{\mathrm{PO}}_{j\_\mathrm{io}}*(1-{\mathrm{IO}}_j)\\ +{\mathrm{\δ}}_{\mathrm{ik}}*{\mathrm{PO}}_{j\_\mathrm{network}}*(1-{\mathrm{Network}}_j)\end{array}$

Wherein α _ik, β _ik, γ _ik, δ _ikrepresent service f respectively _ito CPU, internal memory, hard disk and network four different weights of giving of class resource, and the weights that the different Service Instance of same service is given this four classes resource are identical.Load-balanced server adopts the Web request of industry widely used Weighted Round Robin distribution respective service according to the weights size of each Service Instance.

Provided by the present inventionly ask load-balancing method to be made up of one group of functional module based on the virtualized Web of operating system, they comprise: static information collection module, Dynamic Information Gathering module, weight computing module and load distribution module.

Static information collection module collects the static information of server and Service Instance when load-balancing method initialization.The static information of server mainly comprises: the CPU computing capability P that in Web cluster, every station server is total _{j_cpu}, internal memory disposal ability P _{j_memory}, hard disk I/0 ability P _{j_io}, network throughput capability P _{j_network}.The static information of Service Instance mainly comprises: service f _iservice Instance f _ikplace main frame service f _iservice Instance f _ikcPU weights cpu_share _ik, service f _iservice Instance f _ikmaximum can committed memory memory _{ik_total}.When server in Web cluster or service are dynamically changed, in static information template, need the static information upgrading respective server or service.

Dynamic Information Gathering module mainly collects the state information of server and Service Instance dynamic change.Every T cycle time, Dynamic Information Gathering module can collect the state information of server and Service Instance dynamic change.Cause network burden to prevent information too frequent, cycle time, the value of T was set as 5-10 second according to statistics experience.In addition, when Service Instance is newly-increased or delete, Dynamic Information Gathering module also can perform collection action.The state information of collecting mainly comprises this cycle time of T server s _jthe cpu busy percentage CPU of (1≤j≤n) _j, hard disk I/O load IO _j, offered load Network _j, server s _jthe CPU weights sum of upper all Service Instance copies server s _jupper service f _iexample copy f _ikmemory usage

Weight computing module calculates the final weights of each Service Instance according to the information of collected server and Service Instance.

Load distribution module is according to the Service Instance weights of weight computing module gained, adopt Weight Round Robin algorithm, by more Web Requests routing to the larger Service Instance of weights, less service request is distributed to the less Service Instance of weights, reaches the load balancing between the different Service Instance of same service.

The method that the present invention proposes utilizes the feature that can realize the distribution of runtime system resource elasticity based on the virtualized Docker container of operating system, namely be proportionately distributed to the Docker container of carrying respective service example according to the large young pathbreaker's computational resource of Service Instance CPU weights, realize the load balancing between different Service Instance copy and server.The method by introducing the CPU weights of Service Instance further in Service Instance weight computing formula, thus avoids the situation that the Docker container that same server carries different Service Instance causes server load overweight to the harmful competition of place server residue computational resource.Compared with conventional method, the method of the invention effectively can avoid the different Service Instance copies on same server, because place server load is lower and receive the phenomenon that a large amount of Web request causes the overload of place server simultaneously, realize the load balancing between Service Instance copy and server, improve the concurrent capability of Web group system, reduce the average response time of request.

Accompanying drawing explanation

Fig. 1 is server in cluster, service, Service Instance copy relationship exemplary plot;

Fig. 2 is Dynamic Information Gathering flow chart.

Embodiment

Provided by the present inventionly load-balancing method embodiment is asked mainly to divide four steps based on the virtualized Web of operating system:

(1) server resource information and normalization; (2) server and Service Instance Dynamic Information Gathering (3) according to server resource information and periodic harvest to multidate information calculate the final weights of the Service Instance copy of all services; (4) by WRR dispense request.

(1) server resource information and normalization

Static information collection module collects the CPU computing capability P of every station server in Web cluster _{j_cpu}, internal memory disposal ability P _{j_memory}, hard disk I/0 ability P _{j_io}, network throughput capability P _{j_network}four resource informations.And the maximin calculating them is respectively as follows:

P _{max_cpu}＝max(P _{1_cpu},P _{2_cpu},…,P _{j_cpu},…,P _{n_cpu})

P _{min_cpu}＝min(P _{1_cpu},P _{2_cpu},…,P _{j_cpu},…,P _{n_cpu})

P _{max_memory}＝max(P _{1_memory},P _{2_memory},…,P _{j_memory},…,P _{n_memory})

P _{min_memory}＝min(P _{1_memory},P _{2_memory},…,P _{j_memory},…,P _{n_memory})

P _{max_io}＝max(P _{1_io},P _{2_io},…,P _{j_io},…,P _{n_io})

P _{min_io}＝min(P _{1_io},P _{2_io},…,P _{j_io},…,P _{n_io})

P _{max_network}＝max(P _{1_network},P _{2_network},…,P _{j_network},…,P _{n_network})

P _{min_network}＝min(P _{1_network},P _{2_network},…,P _{j_network},…,P _{n_network})

In order to the impact that the otherness eliminating server isomerism and variety classes resource is brought, adopt Max-Min method to server CPU computing capability P _{j_cpu}, internal memory disposal ability P _{j_memory}, hard disk I/0 ability P _{j_io}, network throughput capability P _{j_network}four resource informations are normalized operation, and use PO respectively _{j_cpu}, PO _{j_memory}, PO _{j_io}, PO _{j_network}value corresponding after representing normalization.

CPU computing capability normalized value PO _{j_cpu}:

PO _{j_cpu}＝(P _{j_cpu}-P _{min_cpu})/(P _{max_cpu}-P _{min_cpu})

Internal memory disposal ability normalized value P _{j_memory}:

PO _{j_memory}＝(P _{j_memory}-P _{min_memory})/(P _{max_memory}-P _{min_memory})

Hard disk I/O ability normalized value PO _{j_io}:

PO _{j_io}＝(P _{j_io}-P _{min_io})/(P _{max_io}-P _{min_io})

Network throughput capability normalized value P _{j_network}:

PO _{j_network}＝(P _{j_network}-P _{min_network})/(P _{max_network}-P _{min_network})

(2) server and Service Instance Dynamic Information Gathering

Relation in Web cluster between server, service and Service Instance copy as shown in Figure 1.In FIG, s in Web cluster ₁, s ₂, s ₃three physical machine run f ₁, f ₂, f ₃8 Service Instance copies (f in Fig. 1 of three services ₁₁, f ₁₂, f ₂₁, f ₂₂, f ₂₃, f ₃₁, f ₃₂, f ₃₃).Every T cycle time, eight the Service Instance copies of the Dynamic Information Gathering flow process of Dynamic Information Gathering module according to Fig. 2 to three servers in Fig. 1 and three services carry out Dynamic Information Gathering, are then recalculated the weights of the Service Instance copy of all services by weight computing module.In addition, when Service Instance is newly-increased or delete, Dynamic Information Gathering module also can perform collection action and the weights of calculation services example copy again.

The information spinner of Dynamic Information Gathering module collection will comprise: cycle time T time server s _jthe cpu busy percentage CPU of (1≤j≤n) _j, hard disk I/O load IO _j, offered load Network _j, server s _jthe CPU weights sum of upper all Service Instance copies server s _jupper service f _iexample copy f _ikmemory usage

(3) Service Instance copy weight computing

According to calculating collected server and Service Instance multidate information in the server resource normalized value of gained and step (2) in step (1), calculating part is deployed in server s _jon Service Instance f _ikfinal weights

${\mathrm{Weight}}_{\mathrm{ik}}^j=[{\mathrm{\α}}_{\mathrm{ik}},{\mathrm{\β}}_{\mathrm{ik}},{\mathrm{\γ}}_{\mathrm{ik}},{\mathrm{\δ}}_{\mathrm{ik}}]*\left[\begin{array}{c}{\mathrm{PO}}_{j\_\mathrm{cpu}}*(1-{\mathrm{CPU}}_j)*\frac{{\mathrm{cpu}\_\mathrm{share}}_{\mathrm{ik}}}{{\mathrm{\Σ}}_{\∀f_{\mathrm{ik}}\∈F_i,L_{\mathrm{ik}}^j\∈s_j}{\mathrm{cpu}\_\mathrm{share}}_{\mathrm{ik}}^j}\\ {\mathrm{PO}}_{j\_\mathrm{memory}}*(1-{\mathrm{Memory}}_{\mathrm{ik}}^j)\\ {\mathrm{PO}}_{j\_\mathrm{io}}*(1-{\mathrm{IO}}_j)\\ {\mathrm{PO}}_{j\_\mathrm{network}}*(1-{\mathrm{Network}}_j)\end{array}\right]$

(4) by WRR dispense request

After saying that weight results upgrades according to the Service Instance weights of (3) step gained, load distribution module adopts Weight Round Robin algorithm, the Web of same service request is balancedly distributed to the different Service Instance copies of this service, thus realizes the load balancing between different instances copy and server.

Claims (1)

1. ask load-balancing method based on the virtualized Web of operating system, it is characterized in that the method comprises the steps:

Step (1). set up server resource state information list; In Web cluster, server set S is expressed as:

S＝{s ₁,s ₂,s ₃，…,s _j,…,s _n}

Wherein s _jrepresent a certain server in Web cluster, 1≤j≤n, n represents the server sum that Web cluster comprises; Server s _jtotal resource P _jbe expressed as:

P _j＝(P _{j_cpu},P _{j_memory},P _{j_io},P _{j_network})

Wherein P _{j_cpu}represent server s _jthe computing capability of CPU, P _{j_memory}represent server s _jinternal memory disposal ability, P _{j_io}represent server s _jhard disk I/O ability, P _{j_network}represent server s _jnetwork throughput; Adopt Max-Min method to server s _jtotal resource P _jvalue is normalized, server s _jtotal resource normalized value PO _jbe expressed as:

PO _j＝(PO _{j_cpu},PO _{j_memory},PO _{j_io},PO _{j_network})

Step (2). set up the list of Service Instance resource state information; If the set of service F that Web cluster externally provides is expressed as:

F＝{f ₁,f ₂,f ₃,…,f _i,…,f _m}

Wherein f _irepresent that i-th kind of service that cluster externally provides, 1≤i≤m, m represent the type service sum that cluster provides; Service f _icomprise multiple different Service Instance copy set F _ibe expressed as:

F _i＝{f _i1,f _i2,f _i3,…,f _ik,…,f _il}

Wherein f _ikrepresent service f _iexample copy, 1≤k≤l, l represent service f _iexample copy number; Service Instance f _ikresource state information ST _ikbe expressed as:

${\mathrm{ST}}_{\mathrm{ik}}=(L_{\mathrm{ik}}^{\text{j}},\mathrm{cpu}\_{\mathrm{share}}_{\mathrm{ik}},{\mathrm{memory}}_{\mathrm{ik}\_\mathrm{total}})$

Wherein represent service f _iexample copy f _ikthe main frame s at place _j, cpu_share _ikrepresent service f _iexample copy f _ikcPU weights, memory _{ik_total}represent service f _iexample copy f _ikmaximum can committed memory;

Step (3). every T cycle time, load-balanced server can the load information of Servers-all in this cycle time T in periodic collection Web cluster; Basic load information comprises:

Server s _jcpu busy percentage:

CPU _j=cpu busy time/(cpu busy time+cpu idle time)

Server s _jhard disk I/O load:

IO _j=hard disk IO the rush hour/(hard disk IO rush hour the+hard disk IO free time)

Server s _joffered load:

Network _j=(flowing into output flow in flow+cycle T in cycle T)/(cycle T * P _{j_network})

Server s _jupper all example copy CPU weights number sums:

${\mathrm{\Σ}}_{{\∀f}_{\mathrm{ik}}\∈F_i,L_{\mathrm{ik}}^j\∈s_j}{\mathrm{cpu}\_\mathrm{share}}_{\mathrm{ik}}^j$

Server s _jupper service f _iexample copy f _ikmemory usage:

${\mathrm{Memory}}_{\mathrm{ik}}^j={\mathrm{memory}}_{\mathrm{ik}\_\mathrm{used}}/{\mathrm{memory}}_{\mathrm{ik}\_\mathrm{total}}$

Wherein memory _{ik_used}represent service f _iexample copy f _ikthe memory value used;

Step (4). load-balanced server is according to the server load information of collecting and existing service device and Service Instance state information, and calculating part is deployed in server s _jon Service Instance f _ikfinal weights and upgrade the value information of respective service example in weights list:

$\begin{array}{c}{\mathrm{Weight}}_{\mathrm{ik}}^j={\mathrm{\α}}_{\mathrm{ik}}*{\mathrm{PO}}_{j\_\mathrm{cpu}}*(1-{\mathrm{CPU}}_j)*\frac{{\mathrm{cpu}\_\mathrm{share}}_{\mathrm{ik}}^j}{{\mathrm{\Σ}}_{{\∀f}_{\mathrm{ik}}\∈F_i,L_{\mathrm{ik}}^j\∈s_j}{\mathrm{cpu}\_\mathrm{share}}_{\mathrm{ik}}^j}+\\ {\mathrm{\β}}_{\mathrm{ik}}*{\mathrm{PO}}_{j\_\mathrm{memory}}*(1-{\mathrm{Memory}}_{\mathrm{ik}}^j)+{\mathrm{\γ}}_{\mathrm{ik}}*{\mathrm{PO}}_{j\_\mathrm{io}}*(1-{\mathrm{IO}}_j)\\ +{\mathrm{\δ}}_{\mathrm{ik}}*{\mathrm{PO}}_{j\_\mathrm{network}}*(1-{\mathrm{Network}}_j)\end{array}$

Wherein α _ik, β _ik, γ _ik, δ _ikrepresent service f respectively _ito CPU, internal memory, hard disk and network four different weights of giving of class resource, and the weights that the different Service Instance of same service is given this four classes resource are identical; Load-balanced server adopts the Web request of Weighted Round Robin distribution respective service according to the weights size of each Service Instance.