METHODS FOR INTELLIGENT DATA TRAFFIC STEERING

[0001] This application is a continuation-in-part of U.S. Utility application Ser. No. 14/730,456, filed on Jun. 4, 2015, and entitled “Adaptive Load Balancer And Methods For Intelligent Data Traffic Steering”, and is related to U.S. Provisional Application Ser. No. 62/008,142, filed on Jun. 5, 2014, and entitled “Adaptive Load Balancer Using Software Defined Network (SDN) For Server Data Congestion Avoidance” and U.S. Provisional Application Ser. No. 62/117,722, filed on Feb. 18, 2015, and entitled “Adaptive Load Balancer And Methods For Intelligent Data Traffic Steering”, the disclosure of each of which is incorporated herein by reference and on which priority is hereby claimed.

[0002] 1. Field of the Invention

[0003] The present invention relates to load balancer devices and structure for controlling the operation of a plurality of servers. This invention further relates generally to data communications, and more specifically, to systems and methods to dynamically load balance servers.

[0004] 2. Description of the Related Art

[0005] Network traffic densities are increasing constantly. The adoption of Software Defined Network (SDN) technology is on the rise for the powerful control it offers over network infrastructure. It is a matter of time before it becomes a part of mainstream adoption. There will be a transitional period in which elements of SDN are used alongside traditional networking technologies and newer overlay solutions. As SDN adoption continues, load balancers will play a critical role in providing the required intelligence for flexible and increasingly effective network deployments.

[0006] More specifically, load balancing is a process of allocating a plurality of client sessions among a plurality of server resources. The goals of load balancing will typically include some combination of maximizing throughput, improving resource utilization, and minimizing average response time. In addition, the ability to load balance among multiple servers will increase reliability through redundancy. If one or more servers should prove unreliable, load balancing can direct client sessions to other remaining reliable servers. Load balancing is commonly used to distribute tasks across a pool of web servers according to various scheduling algorithms. An apparatus that performs the load balancing according to a scheduling algorithm is referred to as a “load balancer.”

[0007] One such scheduling algorithm used by a load balancer for apportioning work among a pool of web servers is round-robin scheduling. In round-robin scheduling, sessions are distributed in equal number to each web server in circular order. Although round-robin scheduling equalizes the number of requests sent to each web server, congestion in the underlying data network connections to each of the servers can vary, resulting in unpredictable delays in response time. Thus, although the number of provided requests is equalized among the web servers, the costs are not equally distributed. Consequently, even though each of the web servers is allocated the same number of requests, the work queue for some web servers may grow long while other web servers may have few requests in their respective queues. Because response time is roughly proportional to the number of requests in a queue, the average response time will suffer when the number of queued requests becomes unequally distributed among web servers.

[0008] In traditional networks there is no end-to-end visibility of network paths, and application traffic is not always routed optimally. More specifically, in conventional networks with load balanced servers, including those that are software defined and controlled by an SDN (software defined network) controller, the load balancer looks at the data traffic congestion at the servers and does not take into account the congestion through the network itself. For example, a path through the network, or a network egress port to which a server is connected, may be overloaded with client requests and server responses thereto. The round robin sequencing, or even a weighted round robin sequencing which is based in part on the data traffic capabilities of the individual servers, performed by conventional load balancers, do not factor in the response times or propagation delays through the network or data velocity at the network egress ports.

[0009] An SDN controller generates statistical information from which such data traffic congestion through the network may be derived. However, heretofore, no conventional load balancer taps this statistical information from the SDN controller to provide adaptive load balancing which dynamically adjusts to changing data traffic congestion through the network for intelligent data traffic steering.

[0010] It is an object of the present invention to provide a system and method which alleviate or avoid data congestion experienced by a plurality of servers controlled by a load balancer in conventional systems.

[0011] In accordance with the present invention, the Load Balancer, integrated with the SDN Controller solution, solves the aforementioned problem by making available critical flow-pattern data. In this way, applications can be routed dynamically across the most optimal server and switching infrastructure.

[0012] An adaptive load balancer formed in accordance with the present invention intelligently steers data traffic through a software defined network (SDN) to which the load balancer is operatively coupled. The network has egress ports to which a plurality of servers is connected. The network has an SDN controller which generates statistical information concerning the network. The adaptive load balancer includes a weighted round robin scheduler module which receives client requests and, based on the statistical information generated by the SDN controller, determines a weight to be attributed to each server of the plurality of servers connected to the network. The adaptive load balancer operates in a passive port mode, a passive path mode, an active path mode and an active path with quality of service (QoS) overlay mode.

[0013] The Load Balancer-SDN Controller combined solution of the present invention enables: 1) Application visibility to network; 2) Network data being “pulled” by the load balancer; 3) Adaptive High Availability load balancing; and 4) Dynamic application delivery.

[0014] These and other objects, features and advantages of the present invention will be apparent from the following detailed description of illustrative embodiments thereof, which is to be read in connection with the accompanying drawings.

[0032] The system and method of the present invention provide intelligent traffic steering through an SDN controlled network using what applicant refers to as a “passive port” method, a “passive path” method, an “active path” method and an “automated QoS control” or “QoS overlay” method, each being in accordance with the present invention. Each of these methods of the present invention, and the systems of the present invention for carrying out these methods, will be described in detail.

[0033] “Intelligent traffic steering”, as its name implies, is a method that, through an adaptive load balancer 2 which uses software defined network (SDN) logic or is coupled to an SDN controller 4, can alleviate or avoid data congestion experienced by a plurality of servers 6. Load balancing, using an SDN controller 4, can obtain reliable and accurate data in real-time with regard to the traffic dynamics of the physical network 8. A physical network is an interconnected set of plurality of packetized data switches and computer servers. The switches and servers can be collocated in a single data center or may be widely distributed across a globally connected network. In addition the servers can be connected to one or more than one packetized data switches. By aggregating this with data on the dynamics of the logical network 8 and application that a load balancer 2 (also referred to herein as an Application Delivery Controller, or ADC) controls, one is able to obtain a truly holistic view and make more intelligent traffic forwarding decisions than in a conventional load balanced server system that does not employ SDN logic or does not communicate with an SDN controller 4. Broadly, the system and method of the present invention can elect to steer application specific traffic around congested parts of the network 8 using passive or active avoidance through an interaction with the SDN controller 4, or can preempt congestion by actively arbitrating the traffic, driven by application specific policies that are injected into the network via the SDN controller 4.

[0034] In a conventional load balanced server network, such as shown in FIG. 1 of the drawings, the load balancer forwards incoming traffic (e.g., video data, voice data, and the like) to the most “appropriate” server 6. The system creates an illusion of a single, highly flexible application server that provides a continuous and consistent quality of service (QoS). However, the conventional load balanced server system lacks real time information on network congestion or other disruptions in the packetized data network and may lead to poor allocation of capital and operational resources, that is, servers 6 may have to be added to keep up with the traffic load. Such load balanced server systems do not have the “intelligence” that the adaptive load balancer 2 and method of the present invention provide.

[0035] As mentioned previously, there are four methods and systems for intelligent traffic steering in accordance with the present invention which are described herein, that is, a “passive port” method, a “passive path” method, an “active path” method and a “QoS overlay” or “automated QoS control” method.

[0036] Before describing each intelligent traffic steering method in detail, reference should be had to FIG. 3 of the drawings. Here, a graph is shown, which compares the characteristics of each of the four methods of intelligent traffic steering of the present invention. As can be seen in the graph of FIG. 3, the abscissa represents the degree of network programmability, and the ordinate represents auto-learning from the server network. The passive port method provides a less deterministic method of intelligent traffic steering than certain of the other methods, such as the QoS overlay method, which is a highly deterministic method of intelligent traffic steering. The passive port method does not leverage network programmability, compared to the active path method and the QoS overlay method of the present invention. More specifically, the passive port method provides congestion awareness and read-only traffic steering and congestion avoidance, and relatively low auto-learning from the network 8. This should be compared to the active path method, which provides network programmability and congestion control, and read-write and policy enforcement in intelligent traffic steering. But, the active path method, like the passive port method, provides a relatively low auto-learning from the network 8.

[0037] The passive path method of the present invention, on the other hand, provides a high degree of auto-learning but, like the passive port method, no network programmability. The QoS overlay method provides, by far, the most intelligent traffic steering of all of the systems, with a high auto-learning capability and providing a high degree of network programmability. Each of the four methods of intelligent traffic steering, and the systems for carrying out such methods, will now be described.

[0038] Reference first should be had to FIGS. 11-13 of the drawings for an overview of how an adaptive load balancer 2 formed in accordance with the present invention functions and how the load balancer 2 is connected to a software defined network (SDN) 8, and further how the adaptive load balancer 2 receives statistical information from the SDN controller 4 of the network 8 regarding throughput, error counters, hop count and other information to determine data traffic congestion through the network 8, which information is factored in and used in determining where client requests should be directed through the network 8 to a particular server 6 connected thereto for intelligent traffic steering.

[0039] Certain terms used herein, in reference to FIGS. 11-13, will now be defined. A “client” refers to the system originating the TCP/IP network requests, also referred to herein simply as “client requests”. The “system” is that which is used to control how requests are distributed to the various targets, and is also referred to herein as an ADC (Application Delivery Controller), or more colloquially known as a “load balancer” 2.

[0040] The “target”, shown in FIG. 11, represents the computer system that will respond to the client requests, usually identified by its IP address. The term “target” used herein and shown in FIG. 11 is synonymous with the server or servers 6 shown in FIG. 12 and also referred to herein.

[0041] The term “network fabric” used herein refers to the packetized data network to handle IP packets, and is usually Ethernet based. An “SDN controller” 4 refers to the active elements in the network 8 that are externally controlled by this system. The SDN controller 4 controls the operation of the SDN switches 10 forming part of the network 8 (see FIG. 12), which directs client requests and return path traffic to and from various targets, or servers 6, connected to different “egress ports” 12 of the network 8.

[0042] The “multiplexor” shown in FIG. 11 refers to the core part of the system, or load balancer 2, that dispatches the client requests to the many targets (i.e., servers 6) based on a weighted round robin algorithm within the load balancer. The multiplexor is depicted in FIG. 12 as the “weighted round robin scheduler” 14.

[0043] The “table” 16 shown in FIGS. 11 and 12 is, essentially, a memory that stores the number of available targets (i.e., servers 6) with static and dynamically calculated weights for each. Each target is assigned a static base weight when it is defined on the Load Balancer. Dynamic weights are calculated by combining the statistical information gathered from the SDN controller 4 by the adaptive load balancer 2. The dynamic weight is factored onto the base weight for each target server and entered into the table 16, along with the server information including the IP address and port number.

[0044] The “SDN-adaptive subsystem” 18 refers to a component of the adaptive load balancer 2 of the present invention which populates the dynamic weights in the table 16 based on statistical data collected from the SDN controller 4. The load balancer 2 uses the statistical information received from the SDN controller 4 to optimize data flow through the network 8 to specific targets, or servers 6, based on criteria in the policy database 20. In the active path and QoS overlay methods of the present invention, the SDN-adaptive subsystem 18 is also used to push information to the SDN controller 4 via the North Bound Interface (NBI).

[0045] The “policy database” 20 is a database forming part of the adaptive load balancer 2 of the present invention that contains all the policies. Policies are sets of conditions, constraints, and settings that allow an administrator to designate how the resources should be allocated to meet the demands of the business to optimize performance, increase utilization or reduce cost. For example the factor/criteria could be application Service Level Agreements (SLA). If the administrator has an application requirement for latency to be less than 50 ms, then the policy defined in the database should require the servers with lower response time to have higher weights in the SDN-adaptive subsystem 18.

[0046] In general, and still referring to FIGS. 11-13, a client request (from a client, such as a laptop computer or a PC external to the network 8 but connected thereto through an Internet connection or an Ethernet connection, for example) is received by a virtual port 22 defined by the adaptive load balancer 2 and is provided to a client interaction module 24 within the balancer 2. More specifically, the client interaction module 24 receives the client requests, which are generally IP based and have a destination address and port, protocol information and other information concerning the request. This client request is provided to the weighted round robin scheduler 14 (see FIG. 12) or multiplexor (see FIG. 11). The weighted round robin scheduler 14 is basically an algorithm that acts upon the client requests received by the load balancer 2 and will change (or provide the client request to the server interaction module 26 shown in FIG. 12 to change) the destination address of the client request based on pre-configured weights attributed to the different servers 6 connected to the network 8 and dynamic statistical information received from the SDN controller 4.

[0047] More specifically, the weighted round robin scheduler 14 looks to the memory (the “table” 16) having information concerning the targets or servers 6, the egress ports 12 and certain base weights defined during initial configuration, as well as the statistical information concerning the network 8 that is received from the SDN controller 4. The SDN controller is polled on a periodic basis (e.g., 5 to 60 second intervals) to gather the network statistical information, and the network statistical information is smoothed using a rolling average over a number of polling intervals. The statistic information includes link utilization, error rates and switch hop count, for example. The smoothed network statistical information generates the dynamic weight values which are updated in the memory for use in the weighted round robin algorithm. The base weights which are stored in the memory, or table 16, of the load balancer 2 relates to the capabilities of the various servers 6 connected to the egress ports 12 of the network 8. For example, an administrator could decide to use CPU configuration for a CPU bound workload as a reference when defining the base weight; for example, if serverl has four times as many CPU cores as server2, then define a weight of 4000 to server1 and weight of 1000 to server2.

[0048] For example, if all of the servers 6 connected to the network 8 have the same capability, the round robin scheduler of a conventional load balancer would direct client requests, or packets of client requests, sequentially to target or server 1, then target or server 2, then target or server 3, continually through target or server n, and then return to target or server 1. A weighted round robin scheduler 14, on the other hand, will take into account a weight assigned to each individual server 6 connected to the network 8. This “base weight” is used to adjust the proportion of client requests sent through the network 8 to each of the targets or servers 6 connected thereto.

[0049] In the adaptive load balancer 2 of the present invention, the weighted round robin scheduler 14, or multiplexor, forming part of the load balancer, takes into account when scheduling where client requests should be directed not only the administratively defined weights of the individual servers 6 connected to the network 8 but also the dynamic conditions of the network 8 and the data traffic congestion therethrough, based on the statistical information received from the SDN controller 4. The SDN controller is connected to each of the network switches and is able to access and store statistical information about the current and historical network conditions. Thus, the adaptive load balancer 2 of the present invention may determine, in real time, how busy not only the servers 6 are but also the paths through and egress ports 12 of the network 8.

[0050] This statistical information is acquired from the SDN controller 4 and provides information such as how much link bandwidth is available for each server 6, the throughput of the egress ports 12 of the network 8 to which the targets or servers 6 are connected, and other information concerning the paths through the various multiple switches 10 of the network 8 from which propagation delays and data congestion through the network paths and at the egress ports 12 of the network 8 may be derived. For example, the weighted round robin scheduler 14 may determine, from this dynamic statistical information received from the SDN controller 4 and the pre-configured weights of the targets or servers 6 stored in the memory (the table 16), a running average of the available bandwidth through the network 8, at the egress ports 12 thereof and that of the various servers 6 connected to the network 8.

[0051] In response to the information it receives, the weighted round robin scheduler 14 (or the server interaction module 26 shown in FIG. 12) will change the destination address and port of the client request or packets of client requests to direct the client requests or packets through the network 8 to the various egress ports 12 of the network 8 and to select targets or servers 6 connected to the egress ports 12.

[0052] FIG. 13 illustrates an example of the scheduling table 28 provided by the weighted round robin scheduler 14 of the adaptive load balancer 2 of the present invention. The table 28 shows that a base weight is assigned to each target or server 6 connected to the network 8, based on the capability of each individual server 6, and then an adaptive weight is included for each target or server 6, which represents the network load, or data traffic congestion through the network 8. This adaptive weight is based on the statistical information received from the SDN controller 4. The adaptive load balancer 2 polls the SDN controller 4 periodically and receives dynamic information concerning data traffic congestion through the network 8, and this information is factored into the weight which is assigned to each server 6 or target connected to the network 8. The resulting weight assigned to each target or server 6 is a function of the base weight and the adaptive weight.

[0053] The client requests, or packets of client requests, are provided to the server interaction module 26 of the load balancer 2. Server interaction module 26 is a network interface component that establishes an IP connection to the appropriate server. It is preferably here that the destination addresses of the client requests are changed based on the determinations made by the weighted round robin scheduler 14. The server interaction module 26 provides the client requests, or packets of client requests, to the SDN switch or multiple SDN switches 10 that are controlled by the SDN controller 4, all of which form part of the network 8, which, based on the destination addresses of the client requests, direct the client requests to the various egress ports 12 of the network 8 and to the targets or servers connected thereto. It should be understood that applications make different demands on the network. For example, a typical web browsing application will have higher levels of data being sent from the server or target to the client. A File Upload application will have higher levels of data being sent from the client to the server or target. A video conferencing application will have a symmetric data requirement to and from the server or target but with strict limits on acceptable delay. It should be understood that statistical network data for the forward and return paths, that is, to and from each individual target or server 6, and through the network 8, is monitored by the adaptive load balancer 2 of the present invention through the NBI connection to the SDN controller.

[0054] A general block diagram of a system incorporating an adaptive load balancer 2, which communicates with an SDN controller 4, for carrying out the four methods of intelligent traffic steering of the present invention is shown in FIG. 2 of the drawings. Since the load balancer 2 operates chiefly at the L4-L7 layer of the network 8, it has application visibility to influence the SDN controller 4 such that upper layer application intelligence can be “pushed” to the SDN controller 4 from the load balancer 2, helping it to make better decisions. Furthermore, network statistical information can be “pulled” from the SDN controller 4 across the North Bound Interface (NBI). This allows the load balancer 2 to make better application load balancing decisions by aggregating its native application intelligence with the information provided by the SDN controller 4. Application intelligence is focused around the health status and performance requirements of the application, including but not limited to the responsiveness of the server or target, application availability and user experience. The methods of the present invention described herein provide a robust mapping between logical and physical networks, and uses the North Bound API (Application Program Interface) to enrich the application information base (AIB) with performance and health characteristics (i.e., such as how many packets of data have been corrupted or dropped) of the physical network 8. The methods of the present invention provide better traffic forwarding decisions.

[0055] As shown in FIG. 2 of the drawings, the adaptive load balancer 2 of the present invention, communicating with the SDN controller 4, directs the SDN controller 4 to activate the SDN switches 10 to direct traffic to one or more servers 6, and appears as a virtual service 22 to the application consumers connected thereto. The solid lines in the block diagram of FIG. 2 show how data is steered to and from the servers 6 by the SDN controller 4 controlling (shown by dashed lines) the SDN switches 10, with input from the adaptive load balancer 2 of the present invention communicating therewith.

[0056] One form of the system of the present invention is shown in FIG. 10. The principles of SDN are focused on the lower layers of the network 8, and load balancers operate chiefly at L4-L7. This provides load balancers application visibility to bridge the gap that exists between the application and the network 8 to influence the SDN controller 4. Upper layer intelligence can be “pushed” to the SDN controller 4 to program flows from the load balancer 2, helping it to improve application delivery across the SDN network.

[0057] Inversely, statistical network data can be “pulled” from the SDN controller 4 across the North Bound Interface (NBI). This allows the load balancer 2 to make better application load balancing decisions by aggregating its native application intelligence with the information provided by the SDN controller 4. The solution of the present invention focuses on the latter as a first step to SDN adaptive load balancing.

[0058] An important augmentation benefit of the load balancer-SDN controller combined solution of the present invention is to improve performance of new and existing applications across the existing SDN network infrastructure.

[0059] For the whole SDN network, the SDN controller 4 has direct access to the lower layer statistical network data including flow metrics that the load balancer does not have access to. The load balancer has visibility of the upper layer application-level information (request load time, SSL TPS, application response throughput, etc.). The load balancer 2 pulls information across the NBI, extending its visibility by adding the statistical network data received from the SDN controller 4.

[0060] In a non-SDN environment, the load balancer is a layer 4-7 device and has no visibility of the lower layer conditions of the individual switching infrastructure. In the SDN environment, the load balancer 2 pulls network statistical information from the SDN controller 4 and utilizes it to make more intelligent load balancing decisions on behalf of clients attempting to access application services. By pulling that information from the SDN controller 4, the load balancer 2 has much better information to make load balancing decisions.

[0061] A block diagram of a system for carrying out the “passive port” method of intelligent traffic steering in accordance with the present invention is shown in FIG. 4 of the drawings. Basically, the congestion level of the port 12 closest to the server 6 negatively impacts the decision to forward the next connection to that server 6. That is, if a port 12 is saturated, the adaptive load balancer 2 will reduce the traffic to it.

[0062] As can be seen from FIG. 4, the adaptive load balancer 2 provides control input to the SDN controller 4 which, in turn, controls the network switches 10 coupled to the servers 6. The blocks shown in the diagram between the switches 10 may be embodied as intelligent IT structures 30, such as firewalls (FW), Intrusion Prevention System (IPS), Spam control, and Application Delivery Controllers (ADC), also known as load balancers. It should be realized, of course, that the adaptive load balancer 2 may provide input to multiple SDN controllers 4 in the server network 8.

[0063] Reference should now be had to FIG. 14 of the drawings, which shows a flow chart illustrating the operation of the system of the present invention in accordance with the “passive port” method. In the passive port method, a user session request arrives at the virtual service interface 22 of the system (Step 1). The Client Interaction Module 24 of the system identifies the request source and potential targets (i.e., servers 6) based on configured parameters, such as the destination application information, the 5 tuple information and the subscriber or tenant ID information (Step 2).

[0064] For all possible targets (i.e., servers 6), the SDN adaptive sub-system periodically polls the SDN controller 4 for network statistical information of the target or server connected ports 12 (i.e., the network ports) via the SDN controller's North Bound Interface (NBI) (Step 3). Then, the SDN adaptive sub-system checks if the SDN controller 4 responds within a specific time-out (Step 4).

[0065] If the SDN controller 4 did respond, then the SDN adaptive sub-system evaluates the statistical information for each target or server connected port 12 for congestion conditions (Step 5). Evaluation is made based on changes in the rolling average from the previous responses from the SDN Controller. The system then checks to see if congestion exists (Step 6).

[0066] If port congestion exists, then the system degrades or changes the weight for that specific target or server 6 connected to the target or server connected port 12 (Step 7).

[0067] The weighted round robin scheduler then identifies the target or server 6 andsends session requests to the application service on the target or server 6 with the best weight (Step 8). The server interaction module sends the session request to the selected target or server 6. The target or server 6 responds through the system (due to source-nat, or NAT), and the system then forwards the response toward the user (Step 9). The session is then established end-to-end through the data path of the network 8 determined by the adaptive load balancer 2 of the present invention (Step 10).

[0068] FIG. 5 is a block diagram of an adaptive load balancer, SDN-controlled server network for carrying out the “passive path” method of intelligent traffic steering in accordance with the present invention. As mentioned previously with respect to a description of the “passive port” method, the congestion level of the network port leading to a server 6 negatively impacts the decision to forward the next connection to that server 6. Hence, a server attached to a congested port receives less traffic. With the “passive path” method of the present invention, the adaptive load balancer 2 accesses statistical network data from the SDN controller for each hop of each path to and from each target or server 6. Statistical data from each switch hop in the network is combined depending on the needs of the application policy. Most commonly this will be done by adding the data from each switch hop to create a total number of switch hops for the complete path. The data for all paths is generated and then compared to the application policy. In the case of an application requiring the shortest path, the numbers of switch hops for each path is compared and the path with the lowest number of switch hops is selected as the optimum path for that user session.

[0069] Again, the blocks between different layers of switches represent intelligent IT structure 30, such as firewalls, IPS, Spam control and ADC (load balancers). The edge switch 32 is connected to network resources external to the SDN controlled network which may include clients accessing resources available on the target or servers 6.

[0070] Reference should now be had to FIG. 15 of the drawings, which shows a flow chart illustrating the operation of the system of the present invention in accordance with the “passive path” method. As shown in FIG. 15, and in accordance with the passive path method of the present invention, a user session request arrives at the virtual service interface 22 of the system (Step 1). The Client Interaction Module 24 of the load balancer 2 identifies the request source and potential targets (i.e., servers) based on configured parameters, such as destination application information, 5 tuple information and subscriber or tenant ID information (Step 2).

[0071] For all possible targets (servers 6), the system periodically polls the SDN controller 4 for statistical network data information for each complete path to each target and statistical information concerning all of the switch ports (including the egress ports 12 of the network 8) via the SDN controller's North Bound Interface (NBI) (Step 3). Then, the system checks if the SDN controller 4 responds within a specific time-out (Step 4).

[0072] If the SDN controller 4 did respond, then the SDN adaptive sub-system 18 of the load balancer 2 compares the calculated values for each of the pre-configured flow paths that are defined by the SDN controller 4 through the network 8 based on the Application requirements in the Policy database, such as the presence of congestion (Step 5). The pre-configured paths are paths programmed by the SDN controller to specify the data flow in both directions from the adaptive load balancer 2 to the target or server 6 on the network. The SDN adaptive sub-system then checks to see if congestion exists (Step 6). Checks are made based on changes in the rolling average from the previous responses from the SDN Controller.

[0073] If path congestion exists, then the system degrades or modifies the weight attributable to an individual target or server 6 (Step 7). The dynamic weights are updated in the memory (table 16). The system then sends the user session request to the application service on the target or server 6 selected by the Weighted Round Robin scheduler (Step 8).

[0074] The target, or server 6, responds through the system (due to source-nat, or NAT) and the system forwards the response toward the user (Step 9). The session is then established end-to-end through the data path of the network 8 determined by the adaptive load balancer 2 of the present invention. (Step 10).

[0075] A block diagram of a server system having an adaptive load balancer 2 which controls the operation of one or more SDN controllers 4 for carrying out the “active path” method of the present invention is shown in FIG. 6 of the drawings. In situations where multiple (redundant) paths lead to the same server 6, the adaptive load balancer 2 programs the SDN controller 4 to take the most effective route through the server system by installing the flows between the load balancer 2 and the selected server 6. For example, and as illustrated by FIG. 6 of the drawings, there may be multiple paths to get from Server B to edge switch 32. The adaptive load balancer 2 looks at the statistical network information and error metrics (e.g malformed packets) at the ingress and egress ports for each hop in each possible forward and reverse path through the switches 10 in the server network 8, which indicate the performance of the paths through the network 8 from the server 6 to the edge switch, and dynamically determines the optimal path to take from the server to the edge switch 32, and directs the SDN controller 4, or controllers, to control the switches 10 through the server network 8 to program the selected path via the SDN Controller. Optimal paths are based on the Application characteristics which may require shortest path, lowest latency path and lowest cost path to be used to deliver on business needs.

[0076] Reference should now be had to FIG. 16 of the drawings, which shows a flow chart illustrating the operation of the system of the present invention in accordance with the “active path” method.

[0077] As shown in FIG. 16, and in accordance with the active path method of the present invention, a user session request arrives at the virtual service interface 22 of the system (Step 1). The system identifies the request source and potential targets (i.e., servers 6) based on the configured parameters, such as destination application information, 5 tuple information and subscriber or tenant ID information (Step 2).

[0078] Then, for all possible targets (i.e., servers 6), the system periodically polls the SDN controller 4 for data path information, statistical network information of all of the possible paths (including the network egress ports 12), and a list of all physical paths by which each of the switches are interconnected in the network via the controller's North Bound Interface (NBI) (Step 3). The system then checks to see if the SDN controller 4 responds within a specific time-out (Step 4). In the event of the SDN controller not responding, the system will retry. If retries fail, then the system will fall back to the previously determined paths to make load balancing decisions.

[0079] If the SDN controller 4 did respond, then the system generates the set of possible flow paths through the network 8 (Step 5.1). A possible path is determined by the ability to reach the address of the target server through the network topology.

[0080] The system then prunes the list of paths based on the application policy database stored in the adaptive load balancer 2 (step 5.2). For example, a path with a large number of switch hops will be eliminated from the list if the application policy database stipulates lowest switch hop count.

[0081] The system selects the new optimal path (Step 5.3) based on the application policy requirements.

[0082] For the optimal path selected, the system sends a data path change command to the SDN controller 4 via the NBI with new data path information to program the alternative path through the network 8 (Step 6).

[0083] The system then sends the session request to the application service on the selected target or server 6 using the optimum path (Step 7).

[0084] The target or server 6 responds through the system (due to source-nat, or NAT), and the system forwards the response toward the user (Step 8). The session is then established end-to-end through the established data path determined by the adaptive load balancer 2 of the present invention (Step 9).

[0085] The “QoS overlay” or “automated QoS control” method of the present invention will now be described, and reference should be had to FIGS. 7 and 8 of the drawings in this regard. First, and as shown in FIG. 7 of the drawings, the QoS overlay method can build on the “active path” or “passive path” method with a QoS capability that will be expressed as an attribute in the policy database for the application (virtual service). In one preferred embodiment, each virtual service can be assigned an ingress and egress traffic classification attribute (eight are shown in FIG. 7, each having a specific priority for various types of traffic specified by the IEEE in the 802.1 standard). The benefit of such a method is an end-to-end, application-centric QoS control of the series and layers of switches 10 of the server network 8. The use of QoS control attributes has the additional benefit of being able to be utilized by series and layers of switches which are beyond the edge switch 32. What is exemplified by FIG. 7 is how layers of switches 10 may be controlled by the adaptive load balancer-SDN controller combination, depending upon the classification of the traffic and the priority assigned thereto.

[0086] For example, background traffic having the least priority (priority “0”) may demand the least amount of control of the various layers of server switches 10, whereas traffic deemed network control, having the highest priority (priority “7”) may be of such an important nature that the adaptive load balancer-SDN controller combination takes steps to insure that the most efficient network flow path through multiple layers of switches 10 to a server 6 is provided. It should be noted herein that the traffic classifications preferably used in the QoS overlay method of the present invention is in accordance with IEEE Standard 802.1P, although it is envisioned to be within the scope of the present invention that different application characteristics and priority attributes may be assigned to the data traffic by the adaptive load balancer 2 of the present invention.

[0087] FIG. 8 is a simplified block diagram of a portion of a load balanced server network 8 for implementing the QoS overlay method of intelligent traffic steering of the present invention. Here, and in accordance with the method, the ingress and egress traffic is assigned a QoS marking attribute via the SDN controller based on knowledge the adaptive load balancer 2 has by recognizing the application traffic or its source. Thus, the first step of the method is to identify the flow using the virtual service which is being addressed and mark the flow of traffic at the incoming edge of the network 8 using NBI API calls to the SDN controller.

[0088] The next step in the method of the present invention is, if the QoS is specified for a particular flow because application centric business requirements require special handling, then the method resorts to SDN controller 4 to identify the flow and add the flow to the appropriate input port of the server switch 10. In addition, a meter 34 is also programmed in the input of the server switch to allow a Service Level Agreement to be tracked and enforced. Meters are configured via the common NBI API on the SDN controller. This allows use of the invention across different hardware switch platforms, which may have differing meter implementations.

[0089] The third step in the “QoS overlay” method of the present invention is to use the programmed meters 34 to take actions when a flow exceeds the Service Level Agreement, drop the packets, or modify the DSCP (Differentiated Services Code Point) which may result in the packets using an alternative lower priority queue on the output port of the server switch 10 and the other switch layers along the end to end path. The SLA characteristics of an Application can include a maximum data rate which an application may use in the SDN controlled network. If an Application exceeds the maximum data rate, then other applications or services may be impacted. The present invention can detect the data rate or bandwidth used by the application, and can take policing actions to protect the bandwidth of other applications. The Application policy database defines the policing actions which can include dropping packets or requests for non-conforming Applications or modifying the QoS policy to use a lower priority queue to reduce the impact of the Application on the end-to-end network. If a lower priority is used, the other applications can be configured to use a higher priority, enabling them to be less impacted by the non-conforming Application. When the data rate has decreased to below the maximum rate, the meter 34 will return to using the original QoS profile.

[0090] Reference should now be had to FIG. 17 of the drawings, which shows a flow chart illustrating the operation of the system of the present invention in accordance with the “QoS overlay or automated QoS control” method. In accordance with the QoS overlay or automated QoS control method of the present invention, the user session request arrives at the virtual service interface 22 of the system (Step 1). The system identifies the request source and potential targets (i.e., servers 6) based on configured parameters, such as destination application information, 5 tuple information and subscriber or tenant ID information (Step 2).

[0091] The system consults the Policy database based on the identified virtual service 22. The policy may define a specific QoS policy for the application which should be applied to the forward and return paths. (Step 3)

[0092] If a QoS policy is present (Step 4), the system will signal the SDN controller via the NBI to program the QoS attributes on the flows at the ingress switch ports on the forward and return paths. If the QoS policy also defines an SLA, this will be programmed via a meter 34 through the SDN controller's NBI. (Step 5)

[0093] The system then sends a session request to the application service on the target, or server 6, (Step 6).

[0094] The target or server 6 responds through the system (due to source-nat, or NAT), and the system forwards the response toward the user (Step 7). The session is then established end-to-end through the data path of the network 8 determined by the adaptive load balancer 2 of the present invention (Step 8).

[0095] FIG. 9 is a diagram which compares a conventional “network centric” system to the “application centric” system and method of the present invention. In the conventional system and method, there may be an excessive use of operational resources (e.g., switches, controllers and servers) to handle the traffic, as the system and method operate on a relatively non-intelligent basis. The “network centric” conventional system applies a deep packet inspection (DPI) to determine the quality of service (QoS) in a non-flexible and non-adaptable manner, categorizing the traffic as being fast, best effort or background, for example. Distinct and separately managed networks may be required to accomplish this, resulting in a costly and over-resourced server network.

[0096] On the other hand, the “application centric” methods and systems of the present invention described herein employ an intelligent converged, QoS server network 8 controlled by an adaptive load balancer/SDN controller combination. The system is entirely automated, dynamic and reliable (since the load balancer 2 has specific knowledge of the Application), and uses a single network fabric with switch independent, open standards and multi-controller architecture. The systems and methods of the present invention intelligently direct traffic through the server network 8 in a cost effective and efficient manner.

[0097] Further features of the adaptive load balancer of the present invention will now be described.

[0098] The adaptive load balancer 2 of the present invention also sets up, and modifies, a table which stores a mapping between each configured target and the weights for each target, and this table is updated when the load balancer 2 is configured with new target server information. Target servers 6 are administratively configured on a load balancer by specifying the IP address and TCP (or UDP) ports associated with the service being offered. Once such servers 6 are configured, the load balancer 2 sends a series of requests to the designated SDN controller 4 responsible for the network 8 to which the target servers 6 are attached. The SDN controller 4 replies to these requests with statistical network data for each path including the network port (“egress port” 12) to which each configured target server 6 is attached. These results are recorded in a table that is later updated during the periodic polling of the SDN controller 4.

[0099] As mentioned previously, the adaptive load balancer 2 periodically polls the SDN controller 4 for statistical information; this dynamically populates the weighted round robin scheduling table 14. Every five seconds, or some other polling interval, the load balancer 2 conducts a poll of the designated SDN controller 4 responsible for the network 8 to which the target servers 6 are attached. Each such poll consists of a series of requests directed to the SDN controller 4 and the compilation of replies from the SDN controller 4 to those requests. Each request seeks the unused network bandwidth available associated with one of the network ports 12 enumerated in the table described above. Each port on a physical network switch has a maximum port speed and therefore maximum bandwidth associated with it based on the type of port. In a modern data center, this would typically be 1 Gigabits per second (Gbps), 10 Gbps or 40 Gbps. For virtual network switches, there is no physical port to define the bandwidth, but it is assigned by the software configuration. For any and all ports, an amount of the bandwidth is being used at any point in time. Ports with high bandwidth utilization have increased congestion or the possibility of congestion. Ports with lower bandwidth utilization at any point in time have a lower possibility of congestion. Large data center networks with many network switches and servers will have connections from clients external to the network connecting to servers inside the network (“North/South traffic”) and additionally connections from servers inside the data center to other servers inside the data center (“East/West traffic”). All types of connections in the data center can introduce congestion which is often very dynamic in nature and hard to detect using traditional load balancers. In relation to the invention described herein, all types of network data can impact server performance and more importantly application performance which this invention mitigates by using SDN adaptive load balancing.

[0100] When a complete set of replies has been received from the SDN controller 4, a calculation is performed to compute new “resulting weight” values for the weighted round robin scheduling table 14. This calculation is performed by multiplying the number representing the base weight of each target server 6 with the number representing the unused network bandwidth associated with the egress port 12 to which the target server 6 is attached. This value is then rounded to the nearest integer and recorded as the dynamic weight in the weighted round robin scheduling table 14. Additionally, the dynamic weight calculated from the data received from the SDN controller is smoothed over a number of polling intervals to reduce rapid fluctuations in the target weights.

[0101] In the event a complete set of replies is not received by the load balancer 2 within, for example, five seconds (or within an alternative administratively determined period of time), the current polling cycle terminates without making any updates to the weighted round robin scheduling table 14, and any partial results received are discarded. Partial results may mean that the connection to the SDN controller is not reliable and the previously calculated results will be used until the next successful polling cycle.

[0102] The adaptive load balancer 2 of the present invention also sets up new client connections, with target server selection. The load balancer 2 enables clients to access applications located on a plurality of target servers 6 through a single interface called a “Virtual Service” 22. A Virtual Service 22 is the single point of access to target servers 6 contact used by clients. Each Virtual Service 22 is uniquely identified by an administratively configured network address, port number, and IP protocol (TCP or UDP).

[0103] When each new client connection request is received at the network interface 22 of the load balancer 2, the load balancer 2 inspects the destination network address, port number, and IP protocol of the request to determine if this set of values matches the configured Virtual Service 22. If no such match can be made, then the request is rejected. If such a match succeeds, then the request is passed to the weighted round robin scheduling module using table 14 to select the best target server 6 for the request.

[0104] By way of example, if Target Servers A, B and C have the resulting weight values of 4, 3, and 2, respectively, in the weighted round robin scheduling table 14, then the next nine valid client requests will have target servers 6 designated in the following sequence: AABABCABC, so long as the values in the weighted round robin scheduling table 14 remain unchanged. The tenth through eighteenth requests will have target servers 6 designated in the same sequence, and so on. The client connections to the plurality of target servers are interleaved within each sequence in relative ratio to the weight values. This is to avoid unintended denial of service attacks on the target servers.

[0105] Once a target server 6 is selected for a client connection request, the request is then passed to the server interaction module 26 in the load balancer 2, where the request is modified for relaying to the target server 6. This is done by replacing the source address and source port information from the client with a source address and source port on the load balancer 2, and by replacing the destination address and destination port information from the client with the address and port of the selected target server 6. An entry is made in a Network Address Translation (NAT) table, recording both the original addresses and the modifications made.

[0106] The connection request is then placed on the network 8 to reach the selected target server 6. Once received by the target server 6, it is processed and a reply is sent to the load balancer 2 at the source address and source port previously modified by the load balancer 2. When the reply is received at the load balancer 2, the server interaction module 26 consults the NAT table to discover the original addressing associated with this reply. It then reverses all previous changes made to the addressing. The reply packet is then placed on a network to reach the requesting client. A validated connection is then established by the client and the server 6.

[0107] Although illustrative embodiments of the present invention have been described herein with reference to the accompanying drawing, it is to be understood that the invention is not limited to those precise embodiments, and that various other changes and modifications may be effected therein by one skilled in the art without departing from the scope or spirit of the invention.