Load Balancing in Recursive Networks using Whatevercast Names

Transcription

1 Load Balancing in Recursive Networks using Whatevercast Names E. Elahi, J. Barron, M. Crotty, Miguel Ponce de Leon and S. Davy {eelahi, jbarron, mcrotty, miguelpdl, Telecommunication Systems and Software Group (TSSG), Waterford Institute of Technology, Ireland Abstract Unlike anycast and multicast, whatevercast comprises of a set of rules which may return one or more results against a name resolution request. The objective of this paper is to highlight how whatevercast name and Name space Manager (NSM) can be exploited for distributed autonomic load balancing and optimal resource utilization inside the RINA (Recursive Internet Architecture) based data center networks. The proposed architecture is based on the feedback principal and runs distributively over the nodes participating in the name resolution process without the need of any dedicated load balancer node. Keywords: Whatevercast name, Recursive Network Architecture, Network Load Balancing I. INTRODUCTION In the last few years, interest has grown in Future Internet architectures [1]. This interest is mainly driven by the pragmatic concerns of large scale ISPs and data center (DC) deployments, cloud providers and businesses that want a more adaptable, configurable, flexible, and resilient network on which to build different services. The fabric of the Internet, its size and scope are the base-ground problems in this space as it is preserved with workarounds which will not meet the requisites needed to build such services [2]. RINA is the most promising architecture with the goal in mind is to provide configurable, secure, resilient, predictable and flexible network services. RINA s communication model is based upon distributed Inter-Process Communication (IPC). As argued in [3] that networking is IPC and only IPC just like communication between two applications over the same host. Two applications on the same host can communicate with each other through local IPC facility provided by the operating system. The same IPC facility can also be extended to allow two applications to communicate on different hosts in the same or in different networks. This extended IPC facility is termed as Distributed IPC Facility (DIF) [2]. Unlike the five-layer model of Internet, in RINA, A DIF can be seen as a single type of layer which can be recursively repeated as many times as required providing the same functions/mechanisms but tuned under different policies to operate over different ranges of the performance space (e.g. capacity, delay, loss). Like the Domain Name system (DNS), there is a directory like service in RINA called Inter-DIF Directory (IDD) or *Work towards this paper was partially funded by the Commission of the European Union, FP7 Collaborative PRISTINE Project, and Waterford Institute of Technology (WIT), Ireland DIF allocator. It is a distributed application that contains the names of the applications and the DIF name from which this application can be accessed. Alongside IDD, there is Name Space Manager (NSM) which keeps the information of application names and the number of instances of each application currently running and how the client applications can access them. When a client application needs to connect with a server application, it will initiate an allocate request to its local IPC Manager (IPCM) and if the local IPCM does not know about the DIF name this server application is enrolled to, then it will send a whatevercast name request to the IDD. The IDD will look for the current instances of the server application and their respective DIFs in consultation with the NSM and reply back to the requesting application. If the IDD does not have the information about the requested application then it will forward the query to its neighbor IDD. Detailed procedure of whatevercast name resolution, NSM and IDD can be seen in the next section. Due to the non-existence of inherent QoS mechanisms, support for multi-homing, self-orchestration and security frameworks in the current TCP/IP based Internet architecture, the data center networks seems to be suffering the most and can be leveraged by employing RINA architecture in data center fabric. Although, in order to provision the support for QoS, security, and multi-homing, the TCP/IP protocol suit is being patched with number of additional protocols and mechanisms such as Diffserv, IPSec, Mobile-IP etc. However each of the additional protocol comes with more additional issues such as scalability, performance, adoptability etc. Optimal resource utilization and load balancing among hundreds of thousands of servers in a data center is imperative to make it energy efficient and to keep the operational costs at minimum level. Presently, dedicated nodes called load balancers are deployed in data centres in order to achieve energy efficient load balancing and optimal resource utilization thus increasing the installation and operational cost of data centres [8]. Support for multi-homing, security and QoS is implicit in RINA. Load balancing in RINA based data centres does not need to have special hardware or dedicated nodes thus minimizing the installation and operational cost of the data centres. In this paper we propose a design for load balancing in RINA based data centers by exploiting whatevercast name and Name Space Manager (NSM) in RINA. Whatevercast refers to one or more members from a set of names. A Distributed Application Facility for Load Balancing (LB-

2 DAF) have been proposed which is a distributed application like IDD that needs to be running in every node along with IDD and NSM. A DAF is a collection of Distributed Application Processes (DAPs) running on different nodes in the network. Each DAP of a DAF has to be registered with one or more common DIFs in order to communicate with each other. The LB-DAF can monitor the current load on the server application to share this information with its neighboring LB-DAF processes which can be termed as Distributed Application Processes for Load Balancing (LB- DAP). The rest of the paper is organized as follows: The next section briefly introduces the process of application discovery in RINA, whatevercast name, NSM and DIF Allocator (IDD). Section III presents the load balancing mechanism in TCP/IP based data centers. Section IV describes how whatevercast name and NSM can be used for load balancing and how this framework can be implemented. In Section V, we discuss the future plan about implementation and thorough testing of the proposed LB-DAF within data center scenario and conclude the paper in Section VI. II. APPLICATION DISCOVERY IN RINA Two processes in a single system discover themselves using the port numbers assigned to them by the operating system. These port numbers are unique addresses in that single system. The operating system provides an IPC facility to let processes communicate with each other. RINA is built with that concept in mind and visualizes the whole Internet as an operating system that provides Distributed IPC facilities (DIFs) to let two or more processes communicate. There may be more than one DIF in the network and each process needs to enroll with one or more DIFs in order to communicate with other processes available in those DIFs. Each DIF has a unique identifier and processes registered in that DIF have a unique name within that DIF [4]. Each process is associated with an instance. There could be two or more instances of the same process running in the network that can be enrolled with the same or different DIFs. Therefore a process can be identified and discovered by using its name, its instance number and the DIF on which it is enrolled. If a client process wants to connect with a server process then it needs to know the name and instance of the server application as well as the name of the DIF on which the server application is enrolled [5]. RINA introduced Whatevercast name and Inter-DIF Directory (IDD) when the client application has no information about the server instance and its relevant DIF. A. Whatevercast Name and Inter-DIF Directory (IDD) When the client application needs to connect with a server application and if it is unaware of the number of instances of the server application and the name of the DIF it is enrolled with, then it will generate and transmit a whatevercast name query. Whatevercast name is an identifier with some associated rules to select the names from a given set [5]. Whatevercast refers to one or more members from a set of names. Multicast and Anycast are special cases of Fig. 1. An example topology: Assuming Link State Routing is used, then all the IDDs in this network should have the information about all the instances of however the IDD at Node-1 should be the first node near source node to know about the existence of all these instances. Whatevercast name. Multicast refers to all the members in a particular set while anycast refers to a single member from a set. The Whatevercast name query which contains the name of the server application process is forwarded to the neighbor nodes. The Name Space Manager (NSM) at the neighbour nodes looks for the process instances currently running and the DIF allocator or Inter-DIF Directory (IDD) looks for the DIF on which these instances are enrolled. Inter-DIF Directory (IDD) [4] is a distributed application which is part of the RINA implementation and run on every RINA node. IDD provides mappings between application process names and their relevant DIFs. Unlike DNS in current Internet architecture, the IDD authenticates the requesting application process to access the requested process. IDD maintains Search Table, Directory Table, Naming Information (using NSM) and Neighbor Table [4]. When the NSM in a particular node receives a Whatevercast name resolution request, it performs lookup for the number of currently running instances of the requested application process from its own name and neighbour tables. If it finds one or more instances for that application process then it checks with the help of IDD for the relevant DIFs on which these instances are registered. After getting the information about the process instances and their relevant DIFs, it replies back to the requested IPC manager (IPCM) of the client node with the number of instances and DIF names otherwise it forwards the same query to its neighboring NSM for resolution. The process of application discovery is illustrated in Figure 2. When the requested IPCM receives a reply from an IDD, it will do one of the three options: 1.) send a resource allocation request to the DIF of the destination application process instance if it is already enrolled on the same DIF; 2.) attempts to join the relevant DIF using DIF extension procedure [5]; 3.) or attempts create a new DIF from source to destination using DIF allocator. An example topology for IDDs can be seen in Figure 1 in order to demonstrate how Whatevercast name resolution works. Let us say that we have a server application pro-

3 Fig. 2. Discovery of the application: Forwarding of the request between the peer IDDs until the destination application is found or the pre-defined termination condition (e.g. TTL) is met. Adopted from Figure by [4] cess named and there are three instances of the same process running at nodes 5, 7 and 20 as shown. Now if the client application process at source node wanted to connect with a particular instance say for example then the IPCM at this node will lookup only for the DIF name on which application process instance is enrolled so that the client can join that DIF or create a new DIF. On the other hand if the client application does not have its preferences for the server application instance then its IPCM will generate a Whatevercast name resolution request for Like the DNS in the current Internet architecture, this query will be resolved according to the node location and the nearest possible availability of the server instance. As can be seen from the given Figure that IDD at node 1 should be the first node near source node to know about all the three instances of the example server application process therefore when the source node transmits a Whatevercast name query to nodes 1 and 31 then node 1 is able to provide the most recent information as it has the shortest route to any of the three instances of Now if the directory table and naming information maintained by each IDD and NSM respectively are extended in such a way that alongwith the DIF name and instance information these tables also contain the information about the load on each server instance then this extended IDD and NSM can be effectively used for balancing the load on each server instance. Section III explains how the IDD can be extended for load balancing purpose. III. LOAD BALANCING IN DATA CENTRES In order for balancing the load between servers in a data centre, an additional entity/node is being used which is called Load Balancer (LBR) [7]. Like NAT, the LBR has one or more public routable IP addresses called virtual IP addresses (VIP) and has one or more servers behind it. The limitation behind this model is that the servers and LBR need to be in the same layer 2 domain. If one or more servers are not in the same layer 2 domain, then such servers would not be able to see the addresses of clients they should be connected to [8]. Therefore, in order for LBR to connect with a server in other layer 2 domains, the packets have to pass through a layer 3 node/router. RINA architecture does not have this limitation. In RINA, servers can be placed anywhere. Application names are location and layer independent; therefore servers can always see the client s application name. The capital cost for networking in data centres is primarily concentrated around the switches, routers and load balancers [8]. Reducing the number of switches, routers and/or load balancers can significantly reduce the capital cost for data centre deployment as well as operational cost. Introducing additional standalone intermediate nodes such as LBR in the end-to-end path might create some performance degradation specifically towards the delay and loss experienced by traffic flows, possibly due to excessive processing load at the LBR. Moreover, in order to avoid a single point of failure and to further balance the load, redundant/additional LBRs are normally deployed in the data centres. Thus making load balancing solution more costly and difficult to maintain. Unlike current Internet architectures, load balancing in RINA based data centres is envisaged to be implemented at the DAF level, rather than deploying additional node/s. DAF based load balancing will utilise a distributed application facility operating at various nodes on the network, which will coordinate with the resources and can redirect network traffic towards lightly loaded servers to make efficient use of resources. Most of the research in load balancing domain [9][10][11][12][13] is centered around the introduction of more and more efficient and intelligent algorithms to run on LBRs in order to balance the load on servers in data center. The algorithms being proposed need dedicated nodes (load balancers) inside data centers. On the other hand, the proposed RINA based LB-DAF is not limited to a single data centre. It is distributed over the network nodes and does not need any dedicated node to do its job thus reducing the capital and operational cost. There will not be any single point of failure. The RINA based LB-DAF provides a load balancing architecture and framework. It is capable to adopt any off-the-shelf load distribution algorithm to achieve the high level business objectives. IV. DAF BASED LOAD BALANCING Load balancing in RINA should enable applications to connect to the most lightly loaded server. In order to do that, we propose an implementation of LB-DAF (Load Balancing Distributed Application Facility) which would be running at every RINA node where IDD service is available. The processes, which are part of this LB-DAF and running on IDD nodes can be referred to as LB-DAP (Load Balancing Distributed Application Process). Each LB-DAP keeps track of the number of server applications running on the system, monitor their current load and share this load information with peer LB-DAPs whenever a change in current load occurs. An abstract implementation architecture for LB-DAF is shown in Figure 3 Each LB-DAP has connectivity to the local NSM and can update the IDD tables. When an LB-DAP receives an update about the from its peers, it instantly updates the local NSM and IDD tables about the server application instances and the load on each instance and forwards an update message to it s neighboring peers. Upon receiving an allocate request from the client application the IPCM will initiate a request to NSM

4 Fig. 3. An abstract architecture of the proposed LB-DAF Fig. 4. Preliminary testing scenario for the proposed load balancing mechanism for the Whatevercast name resolution. The modified NSM now has the information about the server instances as well as the load on each instance, therefore the NSM will reply the requesting IPCM with naming information about the most lightly loaded server instance. And with the help of IDD, the IPCM will also get and enroll with the DIF on which this server instance is currently registered and available. A. Implementation and Experimentation Preliminary test for load balancing was conducted using three virtual machines having RINA prototype [6] installed. Two server instances of a file transfer server application were made available on two different VMs. Each VM was enrolled on a common DIF named as normal DIF which was constituted by three IPC processes Test1.IRATI, Test2.IRATI and test3.irati each running on different VM as shown in Figure 4. The client VM is connected with each server VM through distinct non-overlapping virtual LAN clusters VLAN1 and VLAN2. There is a Shim DIF created by Shim IPC Processes to facilitate the hop-by-hop communication between two or more hosts in a single VLAN as shown. It is assumed that the client is aware of the both server instances available through Normal DIF. Therefore instead of sending an allocate request with Whatevercast name, the client sends an allocate request to the IPCM by specifying the server application name along with its instance. In this way the client can connect with more than one server if it knows the number of instances already running thus aggregating the available bandwidth on different interfaces. The experiment was conducted by transferring 20GB file placed on both server instances and 50% gain was achieved when the file was downloaded in parallel from both instances as compared to downloading from a single instance and each server has to transfer 50% less amount of data with half connection lifetime. There are two aspects to consider for load balancing in RINA: 1. Selection of server instance to connect to if multiple clients contend for the same server. If a client does not give any server preference, e.g. it just wanted to access example.com, then the NSM will decide which server instance to connect to and allocate a flow. By using this approach the NSM has a better view of allocations and could balance the load at servers and eventually clients can experience better throughput. This can be achieved by the proposed LB-DAF. 2. Re-ordering of received packets if a client connects to multiple servers and out of ordered data packets from the servers are received by the client. This is the case when there are multiple servers for the same service under a single administrative domain e.g. and etc. The client application process can choose the server/s to connect to. For example, if there are two file servers having a specific file of size 2GB. The client may connect to both the servers and request half of the file from server 1 and the other half from server 2. The client application needs to know the reordering of data packets if it intends to connect with multiple servers in order to aggregate the bandwidth available at it s multiple interfaces. The client must keep track of all the data being received from each server because if the connection with one server lost then the client application should be able to receive all the remaining data from the other connection/s. This approach could reduce the load on servers, enhance the throughput and aggregate the bandwidth if each flow adopt distinct path. Because, if the flows pass through a common intermediate node, then the available capabilities at that node need to be shared among each flow that might cause performance degradation. One more point to note is that if the client application knows the number of server instances then it can choose one instance and initiate a connection with it. However this server instance may not necessarily be the most lightly loaded server because the client application has no way to check the load on server instances. Therefore getting load statistics of each server instance through LB-DAF and then initiate connection with the most suitable server instance could produce desired network performance. V. FUTURE WORK In order to test the performance of LB-DAF in a large network and data center scenario, we have planned to use the virtual wall test bed provided by Fed4Fire [14]. The virtual wall test bed provided an emulation environment of more than 100 nodes of dual processor and dual core servers. All the nodes are connected with each other through a switch

5 Fig. 5. Virtual Wall testbed topology provided by Fed4Fire. Figure adopted from [14] Fig. 6. An example topology to test LB-DAF on virtual wall testbed. The the dotted filled circles shows the client nodes. The solid filled circles represents the nodes where IDD and LB-DAF processes are running while the unfilled circles represents the RINA based forwarding nodes. Such nodes will not participate in load balancing process however they will participate in routing process. The solid lines show the active connections and at least the presence of one DIF between two inter-connecting nodes while the broken lines show in-active connection and the absence of any DIF. having 1.5Tbps of non-blocking switching back-plane capacity as shown in Figure 5. The nodes are configurable and customizable to create and emulate any network topology. The topology proposed for LB-DAF testing can be seen in Figure 6. There will be different server farms virtually located on different geographical locations. Instances of a file server will be running on different server machines in one or more server farms. The arbitrarily spread client nodes will initiate one or more connections with the file server by just specifying the Whaevercast name of the file server. The LB-DAF running on arbitrary nodes will guide the clients to connect to the most lightly loaded server as per load balancing policy. The performance parameters are but not limited to percentage utilization of each server machine, average throughput gained for each flow, average response time and energy consumption by each server machine with and without LB-DAF. VI. CONCLUSION In this paper, Load Balancing Distributed Application Facility (LB-DAF) is presented in order to balance to the load on the RINA based servers in data centres. As the LB-DAF works in distributed manner and updates the Name space Managers (NSM) of its neighbouring nodes about the load on servers therefore it enables the client nodes to initiate the connection to the most lightly loaded server. The LB-DAF could abandon the use of any dedicated nodes in data centres for load balancing without degradation in service quality. Moreover as the LB-DAF updates its peers about the current load on the server instances therefore it is able to facilitate the load balancing among server instances available not only within data centers but also at distributed locations. REFERENCES [1] Investigating RINA as an Alternative to TCP/IP (IRATI), Website [Accessed August 06, 2015]. [2] J. Day: Patterns in Network Architecture: A Return to Fundamentals, Prentice Hall, [3] John Day, Ibrahim Matta, and Karim Mattar, Networking is IPC: a guiding principle to a better internet, In Proceedings of the 2008 ACM CoNEXT Conference (CoNEXT 08). ACM, New York, NY, USA, Article 67, 6 pages. [4] Eleni Trouva, Eduard Grasa, John Day and Steve Bunch, Layer Discovery in RINA networks, IEEE 17th International Workshop on Computer Aided Modeling and Design of Communication Links and Networks (CAMAD), 2012 [5] RINA Specifications, [Accessed on August 14, 2015] [6] RINA Prototype Implementation, [Accessed September 14, 2015] [7] Cisco systems: Data center: Load balancing data center services, 2004 [8] Albert Greenberg et. al.,the Cost of a Cloud: Research Problems in Data Center Networks, Published in ACM SIGCOMM Computer Communication Review archive Volume 39 Issue 1, January 2009 [9] Aameek Singh et. al., Server-Storage Virtualization: Integration and Load Balancing in Data Centers, IEEE International Conference for High Performance Computing, Networking, Storage and Analysis, [10] M. Randles, Lamb, D., Taleb-Bendiab, A., A Comparative Study into Distributed Load Balancing Algorithms for Cloud Computing, IEEE 24th International Conference on Advanced Information Networking and Applications Workshops (WAINA), 2010 [11] Sourabh Bharti, K.K.Pattanaik, Dynamic Distributed Flow Scheduling with Load Balancing for Data Center Networks, The 4th International Conference on Ambient Systems, Networks and Technologies (ANT 2013) [12] Dharmesh Kashyap, Jaydeep Viradiya, A Survey Of Various Load Balancing Algorithms In Cloud Computing, INTERNATIONAL JOUR- NAL OF SCIENTIFIC & TECHNOLOGY RESEARCH (ISSN ) VOLUME 3, ISSUE 11, NOVEMBER 2014 [13] Geethu Gopinath, Shriram K. Vasudevan, An in-depth analysis and study of Load balancing techniques in the cloud computing environment, 2nd International Symposium on Big Data and Cloud Computing (ISBCC15), Procedia Computer Science 50 ( 2015 ) [14] Federation for Future Internet Research and Experimentation (Fed4Fire), Website [Accessed September 13, 2015]