Various Schemes of Load Balancing in Distributed Systems- A Review

Transcription

1 741 Various Schemes of Load Balancing in Distributed Systems- A Review Monika Kushwaha Pranveer Singh Institute of Technology Kanpur, U.P. (208020) U.P.T.U., Lucknow Saurabh Gupta Pranveer Singh Institute of Technology Kanpur, U.P. (208020) U.P.T.U., Lucknow Abstract Over the past three decade, parallel and distributed computing witnessed major growth due to the declined cost of hardware, advancement in technology, explosive growth of internet and need to solve large-scale problems. These systems often face the problem of load imbalance, which can degrade the performance of the system. Load balancing improves the system performance by dividing the work load effectively among the participating computers/nodes. This paper presents various aspects of load balancing and explains various load balancing approaches and strategies and their respective algorithms. Keywords: Load balancing, Thresholding, Migration, Distributed Systems. 1. INTRODUCTION Outstanding progress in computer technology has lead to the rise of the demand of high speed processing and the need of high scalability, availability and rapid response. This resulted in use of parallel and distributed computing systems where more than one processor processes the job simultaneously. One of the main research issues in parallel and distributed system is effective strategy to distribute workload among multiple processors. Load balancing is used for minimizing the response time, maximizing the throughput, and to avoid the overload. Load balancing is to ensure that every processor in the system does approximately the same amount of work at any point of time [2]. 1.1 Load Balancing Load balancing is the process of distribution or redistribution of load among processor therefore improving the performance of the system. In computing, load balancing distributes workload across multiple computing resources, such as computers, a computer cluster, network links, or central processing units. Load balancing aims to optimize resource use, maximize throughput, minimize response time, and avoid overload of any single resource [3] Need of Load Balancing A distributed system contains number of processors working independently with each other and linked by channel. Some are not linked with any channel. Each processor has an initial load that is the amount of work to be performed, and each may have a different processing capacity. The workload has to be evenly distributed among all processors based on their processing speed so that time to execute all tasks gets minimized and idle time of each processor can be reduced. This is why we need load balancing. Load imbalance is also a main problem in data parallel applications and here also it mainly occurs due to the uneven distribution of data among the various processors in the system. Without good load distribution strategies and techniques, we cannot aim to reach good speedup and good efficiency Issues related to Load Balancing The following issues are analyzed during load balancing: a) In distributed environment the channels are of finite bandwidth and the processing units may be physically distant hence load balancing need to take decision of whether to allow task migration or not. b) A computing job may not arbitrarily divisible leading to certain constraints in dividing tasks. c) Each job consists of several smaller tasks and each of those tasks can have different execution times. d) The load on each processor as well as on the network can vary from time to time based on the workload brought about by the users. e) The processors capacity may be different from each other in architecture, operation system, CPU speed, memory size, and available disk space. Taking above factors into consideration the load balancing can be generalized into four basic steps: a) Monitoring processor load and state b) Exchanging load and state information between processors c) Calculating the new work distribution d) Actual data movement

2 Policies of Load Balancing A good load balancing algorithm is defined by some basic policies [10] which are as follows: Information policy: Specifies what workload information to be collected, updated, When it is to be collected and from where. a) Triggering policy: Determines the appropriate period to start a load balancing operation. b) Resource type policy: Retrieve the task from most overloaded resource or not. c) Location policy: Uses the results of the resource type policy to find a suitable partner for a server or receiver d) Selection policy: Defines the tasks which should be migrated from busiest resources (source) to most idle resources (receiver). A selection policy considers several factors in selecting a task, for e.g. transfer of small task will take less overhead Advantages of Load Balancing Some major advantages of load balancing [5, 6] are as follows: a) It reduces the task waiting time. b) It minimizes task response time. c) It maximizes utilization of system resources. d) It maximizes system throughput. e) It improves reliability, and stability of the system. f) It accommodates future modification. g) Long starvation is avoided for small jobs. h) In load balancing overall system performance is enhance by improving the performance of each node. Because of the above benefits, the load balancing strategies become a field of intensive research. As a result, a large number of scheduling and load balancing algorithms have been developed in the past several years. 2. LOAD BALANCING APPROACHES In load balancing there are two approaches and each approach have their own algorithms. The two approaches are as follows: 2.1 Static Load Balancing In the Static load balancing approach, the load balancing decisions are made deterministically or probabilistically at compile time according to the performance of computing nodes and remain constant during runtime. Number of tasks in each node is fixed in this approach [5]. Static load balancing methods are non-preemptive i.e. once the load is allocated to the node it cannot be transferred to another node. 1. It has minimum delays. 2. Algorithms are simple and easy to implement. 3. System overhead is minimized. 1. Task cannot be migrated while execution. 2. Overall performance of system decreased due to load fluctuations. 3. Less useful when task have different execution time and nodes are heterogeneous Algorithms of Static Load Balancing Among various proposed algorithms major static load balancing algorithms are as follows: Round Robin Algorithm Round Robin Algorithm (RRA) assigns tasks sequentially and evenly to all the nodes. All tasks are assigned to computing nodes based on Round Robin order, meaning that computing nodes choosing is performed in series and will be back to the first computing node if the last computing node has been reached [5]. Each node maintains its load index locally independent of allocations from remote node. 1. Inter-process is not required. 2. Useful for jobs of equal processing time and nodes of same capabilities. 1. Not useful when tasks have unequal processing time. 2. Not useful when nodes have different capacities Randomized Algorithm Randomized Algorithm (RA) uses random numbers in selecting computing nodes for processing, without having any information about the current or previous load on the node. The computing nodes are selected randomly following random numbers generated based on a statistic distribution [5, 1]. 1. It works well for particular special purpose applications. 2. No inter process is required. 1. It is not considered elegant solution. 2. Maximum response time among all algorithms Central Manager Algorithm In each step in Central Manager Algorithm (CMA), central node selects slave node to be assigned a task.

3 743 Slave node having least load is being selected. The central node maintains the load index of all slave nodes connected to it. Whenever, load is changed, a message is send by the slave nodes to the central node [2]. The load manager makes load balancing decisions based on the system load information, allowing the best decision when of the process created [5, 2]. 1. Performs well when dynamic activities are created by different hosts. 1. CMA needs high level of inter process. 2. Single point of failure Threshold Algorithm In Threshold Algorithm (TA) tasks/processes are assigned immediately upon creation to the computing nodes (processors). Computing nodes for new processes are selected locally without sending remote messages. Each node keeps a private copy of the system s load information. The load of a computing node can be characterized by one of the three levels which are: under loaded, medium and overloaded. Two threshold parameters t_under and t_upper can be used to describe these levels [5, 7]: 1. Under loaded: workload < t_under, 2. Medium : t_under workload t_upper, 3. Overloaded: workload > t_upper. In the initial stage, all the computing nodes are considered under loaded. Whenever the load state of a computing node exceeds the load level threshold, then it sends messages regarding the new load state to all of the other computing nodes, regularly updating them so that as the actual load state of the entire system can be known to every node. If the local state is not overloaded then the process is allocated locally. Otherwise, a remote under loaded processor is selected, and if no such host exists, the process is also allocated locally. 1. It has low inter process. 2. Performance is improved because of large number of local process allocations which decreases the overhead of remote process allocations and remote memory accesses. Whenever all remote processors are overloaded then all processes are allocated locally causing two major problems:- 1. One overloaded processor may have much higher load than other overloaded processors, causing significant disturbance in load balancing. 2. Increased execution time of an application. 2.2 Dynamic Load Balancing There are three general classes of problems where a static load balancing is either impossible or can lead to imbalance of load, problems are: 1. The first class consists of problems in which all the tasks are available at the beginning of the computation but the amount of time required by each task is different. 2. The second class consists of problems in which tasks are available at the beginning but as the computation progresses, the amount of time required by each task changes. 3. The third class consists of problems in which tasks are not available at the beginning but are generated dynamically. In static load balancing too much information about task and system is required before execution which is not possible every time like in these three classes of problems. So Dynamic load balancing was developed to address these constraints. Dynamic load balancing makes more informative load balancing decisions during execution by the runtime state information [1]. In dynamic load balancing algorithms work load is distributed among the processors at runtime. These algorithms monitor changes on the system work load and redistribute the work accordingly. 1. Dynamic load balancing works well for heterogeneous systems. 2. Task can be redistributed to any processor while run time hence overloading and under loading problems become minimum. 3. It works well for task having different execution time. 4. The system need not be aware of run-time behavior of the applications before execution. 1. High over heads occurs and becomes more when number of processors increase. 2. Dynamic load balancing algorithms are complex hence not very easy to implement. 3. System overhead increases because it is preemptive.

4 Categories of Dynamic Load Balancing In general, dynamic load-balancing algorithms can be broadly categorized as follows: a) Centralized or Decentralized In centralized approach, only one processor acts as the main or central controller (Master). Master node process or holds the collection of tasks to be performed. Master node has global view on the load information of all nodes connected to it, and decides how to allocate jobs to each of the nodes. And rest of the nodes act as slaves [10]. Tasks are sent to the slave processes for execution. It has advantage of ease of implementation, but suffers from single point of failure. In Decentralized approach, all nodes in the system are involved in making the load balancing decision. The state information is spread among all the nodes and nodes are responsible in managing their own resources or allocating tasks residing in their queues to other nodes. It is more scalable and has better fault tolerance. b) Cooperative or Non-cooperative If a distributed load balancing algorithm is adopted, the next issue that should be considered is whether the nodes involved in the task scheduling process are working cooperatively or independently (non-cooperatively). In the non-cooperative approach, each node is independent has autonomy over its own resource scheduling, that is, decisions are made independently of the rest of the system and therefore the node may migrate or allocate tasks based on local performance. In the cooperative scheme, each load balancing decision maker (scheduler) has the responsibility to carry out its own portion of the scheduling task, but all schedulers are working toward a common system wide goal [5]. c) Adaptive or Non-Adaptive Dynamic load balancing is said to be adaptive if scheduled decisions take into consideration past and current system performance and are affected by previous decisions or changes in the environment. It is said to be non-adaptive if parameters used in load balancing remain the same regardless of system's past behavior. d) Sender/Receiver/Symmetrical Initiated Distribution of task in dynamic load balancing can be sender initiated, receiver initiated or symmetrical initiated. In sender initiated strategy, Congested nodes attempt to move work towards under-loaded nodes. In receiver-initiated strategy, under-loaded nodes request tasks to be sent to them from nodes with higher loads. In the symmetric approach, both the under-loaded as well as the overloaded nodes may initiate load transfers. Algorithms of Dynamic Load Balancing There are three basic dynamic load balancing algorithms which are as following: Central Queue Algorithm Central Queue Algorithm [9] stores new activities and unfulfilled requests in a cyclic FIFO queue on the main host. Each new activity arriving at the queue manager is inserted into the queue. Then, whenever a request for an activity is received by the queue manager, it removes the first activity from the queue and sends it to the requester. If there are no ready activities in the queue, the request is buffered, until a new activity is available. When a processor load falls under the threshold, the local load manager sends a request for a new activity to the central load manager. The central load manager answers the request immediately if a ready activity is found in the process-request queue, or queues the request until a new activity arrives. This is a centralized initiated algorithm and need high among nodes Local Queue Algorithm This algorithm supports inter process migration. The main concept in local queue algorithm is static allocation of all new process with process migration initiated by the host when its load falls under the predefined minimum number of ready processes (threshold limit). Initially, new processes created on the main host are allocated on all under loaded hosts. From then on, all the processes created on the main host and all other hosts are allocated locally [9, 2]. When the local host gets under load it request for the activities from the remote hosts. The remote hosts than look up its local list for ready activities and compares the local number of ready activities with the received number. If the former is greater than the latter, then some of the activities are passed on to the requestor host and get the acknowledgement from the host. This is a distributed co-operative algorithms and requires inter-process but lesser as compared to central queue algorithm Least Connection Algorithm This algorithm decides the load distribution on the basis of connections present on a node [13]. The load balancer keeps track of the numbers of connections attached to each node and selects the node with least number of connections for load transfer. The number increases when a new connection is established and decreases when connection finishes or time out. Least Connections methods work best in environments where the nodes have similar capabilities. Load imbalance may be caused when the tasks are all of

5 745 different durations because Connection counting just doesn t account for that scenario very well. a) Nature: It tells whether algorithm is static or dynamic. b) Overhead: It is amount of overheads like interprocess, migration of tasks etc. involved while implementing the algorithm and should be minimum. c) Resource Utilization: It tells whether the algorithm is able to utilize all the resources optimally or not means less idle processors. d) Process Migration: It tells when a system will migrate its process. The algorithm is capable of deciding when it should make changes of load distribution during execution of process or not. e) Fault Tolerant: It tells whether the algorithm can work continuously in event of failure or not, performance of algorithm is decreasing or not. f) Response Time: It is time a distributed system using a particular load balancing algorithm is taking to respond and must be less. g) Waiting Time: It is the time period spent waiting in the ready queue and should be less. h) Centralized or decentralized: It tells whether algorithm is centralized or decentralized. i) Cooperative: It defines the extent of independence that each processor has in concluding that how should it can use its own resources. j) Adaptability: It tells if algorithm can adapt to changing situations. 2.3 Comparison on different Algorithms The comparison of above given algorithms of static and dynamic load balancing considering the following parameters is as follows: Table1: Comparative analysis of load balancing algorithm. D- Decentralized, C- centralized Parameters Algorithms Nature Overhead Resource Utilization Process Migration Fault Tolerant Response Time Waiting Time Centralized/Decentrali zed Adaptability Roun d Robi n Randomize d Central Manage r Threshol d Central Queue Local Queue Least Connecti on Static static Static dynamic dynamic dynamic Dynamic Low low Low high high high High Less less Less less less more More No no No no no yes No No no Yes no yes yes No less less Least less more more Less more more More more less less Less D D D D C C D less less Less less more more More 3. Summary of Load Balancing Algorithms There are many other dynamic and static algorithms. Here we have summarized all the above algorithms as well as some other frequently used algorithms of both approaches.

6 746 Table 2: Summary of Load Balancing Algorithms Approaches Algorithms Key features Advantages Disadvantages Static Round Robin Weighted Round Robin Assigns task sequentially. Chooses computing nodes serially. Tasks assigned according to their weights. No inter-process Simple and easy to implement. Minimum system overhead. Improvement over round robin. Useful for nodes of different capacities. Not useful if tasks have varying execution time. Not useful if nodes have different capacities. Not useful if tasks have different execution time. Randomized Central Manager Threshold Nodes selected randomly. Master node selects slave nodes. Least load node is selected. If load > threshold limit task is assigned. No inter-process. Useful for special purpose applications. Useful when dynamic tasks are created. Minimize remote process allocations. Minimize remote memory accesses. Maximum response time. Uneven load distribution. High inter-process. Bottleneck state arises. Increased execution time. Disturbed load balancing. Dynamic Central Queue Local Queue Least Connection Stores new task in cyclic FIFO queue. Has main host. Receiver initiated balancing. Inter-process task migration if load < threshold limit. Load migrated to node with least connections. Useful for heterogeneous nodes. Tasks assigned after execution. Less inter-process. Useful for nodes of similar capabilities. Single point of failure. No inter process task migration. Not useful for tasks of different durations. Nearest neighbor Task migrated to immediate neighbor nodes. Symmetric Less delay. Decentralized

7 747 Random Probabilistic Threshold Centralized Information and Centralized Decision Centralized Information and Distributed Decision Distributed Information and Distributed Decision initiated load balancing. Task migrated to randomly selected node if load > threshold. Load information not exchanged. Nodes have a load vector. Load vector shared periodically with random nodes. Load information exchanged. Sender initiated load balance. Task is accepted if node has load > threshold. Master node keeps load information. Maser node takes load balancing decisions. Master node keeps load information. Load balancing decision is distributed. Each node takes own load balancing decision. Load information is distributed. nature. Simple and easy to implement. Minimized overhead. Simple implementation All resources are utilized. Efficient performance results. Less inclusion of message information. Robust nature. No single point of failure. Efficient performance. Communication overheads. Poor utilization of system resources. Algorithm is not scalable. More overhead. Single point of failure. Not scalable. Not always up-todate. Greater overhead occurs. More overhead sometime. 4. CONCLUSION In this paper we have taken review of concepts about load balancing and various load balancing strategies. Computing resources are rapidly developing, and growing uses of heterogeneous system with parallel or distributed computing issue of load imbalance has emerged and load balancing is solution to such issues. Here we presented that load balancing distributes the load evenly among nodes, hence increasing overall performance. Then we studied various policies which should be considered while designing the load balancing algorithm like information policy etc. Finally we

8 748 discussed Static and Dynamic load balancing strategies with their respective algorithms and concluded that each strategy has their own pros and cons and there exists no absolutely perfect balancing algorithm but one can use depending on the need. REFERENCES [1] Mohsen and Hossein Delda, Balancing Load in a Computational Grid Applying Adaptive, Intelligent Colonies of Ants, Informatica 32 (2008) [2] Sandeep Sharma, Sarabjit Singh, and Meenakshi Sharma, Performance Analysis of Load Balancing Algorithms, World Academy of Science, Engineering and Technology, [3] mputing%29. [4] [5] S. F. El-Zoghdy and S. Ghoniemy, A Survey of Load Balancing In High-Performance Distributed Computing Systems, International Journal of Advanced Computing Research, Volume 1, [6] Mr. Amitkumar S Manekar, Mr. Mukesh D Poundekar, Prof. Hitesh Gupta and Prof. Malti Nagle, A Pragmatic Study and Analysis of Load Balancing Techniques In Parallel Computing, International Journal of Engineering Research and Application, Vol. 2, Issue4, July-August 2012, pp [7] Daniel Grousa and Anthony T., Non-Cooperative load balancing in distributed systems, Journal of Parallel and Distributing Computing, [8] P. A. Tijare and Dr. P. R. Deshmukh, Schemes for Dynamic Load Balancing A Review, International Journal of Advanced Research in Computer Science and Software Engineering 3(6), June , pp [9] William Leinberger, George Karypis and Vipin Kumar, "Load Balancing Across Near-Homogeneous Multi-Resource Servers", /00, 2000 IEEE. [10] Mr.Gaurav Sharma and Jagjit Kaur Bhatia, A review on different approaches for load balancing in computational grid, Journal of Global Research in Computer Science, Volume 4, No. 4, April [11] Sherihan Abu Elenin, and Masato Kitakami, Performance Analysis of Static Load Balancing in Grid, International Journal of Electrical & Computer Sciences, IJECS-IJENS, Vol: 11, No: 03,june [12] d-balancing-scheduling-methods-explained. [13] Suriya and Prashanth, Review of load balancing in cloud computing, IJCS, vol.10 issue.1, Jan [14] Javed hussain and Durgesh kumar mishra, Analysis and investigation of nearest neighbor algorithm for load balancing, International Journal of Emerging Trends & Technology in Computer Science, Volume 2, Issue 6, November December [15] Sherihan Abu Elenin and Masato Kitakami, Performance Analysis of Static Load Balancing in Grid, International Journal of Electrical & Computer Sciences, Vol: 11, No: 03, June 2011.