Applying Self-Aggregation to Load Balancing: Experimental Results

Transcription

1 Applying Self-Aggregation to Load Balancing: Experimental Results Elisabetta Di Nitto, Daniel J. Dubois, Raffaela Mirandola Dipartimento di Elettronica e Informazione Politecnico di Milano Fabrice Saffre, Richard Tateson BT Group Bionetics 008 Awaji Island, Hyogo, Japan November th, 008

2 Motivation Existing Load-Balancing algorithms are not efficient in heterogeneous highly dynamic settings Examples: Nodes can be small networked devices like sensors; Nodes are not stable : they can move, disappear and come back again (dynamism); They can be specialized, for example to save energy they can elaborate only particular data (heterogeneity). Problem: balance the workload among these nodes in a self-organized way when the setting is heterogeneous and the structure of the network changes over time.

3 Problem and Assumptions Is it possible to balance the workload in a network of interconnected heterogeneous nodes? Consider a set of nodes;

4 Problem and Assumptions Is it possible to balance the workload in a network of interconnected heterogeneous nodes? Consider a set of nodes; Nodes are connected in an overlay network;

5 Problem and Assumptions Is it possible to balance the workload in a network of interconnected heterogeneous nodes? Consider a set of nodes; Nodes are connected in an overlay network; Each node belongs to a type;

6 Problem and Assumptions Is it possible to balance the workload in a network of interconnected heterogeneous nodes? Consider a set of nodes; Nodes are connected in an overlay network; Each node belongs to a type; Each node has a queue of jobs of its type;

7 Problem and Assumptions Is it possible to balance the workload in a network of interconnected heterogeneous nodes? Consider a set of nodes; Nodes are connected in an overlay network; Each node belongs to a type; Each node has a queue of jobs of its type; Each node may receive jobs from the environment of from neighbor nodes;

8 Problem and Assumptions Is it possible to balance the workload in a network of interconnected heterogeneous nodes? Consider a set of nodes; Nodes are connected in an overlay network; Each node belongs to a type; Each node has a queue of jobs of its type; Each node may receive jobs from the environment of from neighbor nodes; Each node may only receive or process jobs of its own type.

9 Peer-to-peer network: No centralized control; Running Scenario No global knowledge over the network; Each node may interact with its neighbors only; High node churn. Heterogeneity: Nodes are specialized workers; Nodes may have different service times when executing jobs. Node interactions: Nodes can act as matchmakers and rewire the network; Nodes can send their jobs to their neighbors (but only if the type of the job matches the type of the node).

10 Our Goal: Load Balancing Balancing the workload of the network among the nodes in such a way to parallelize as many jobs as possible Goal 3 Maximize throughput={# completed jobs}/{elapsed time} Starting point: Bio-inspired Self-Aggregation techniques; Dimension Exchange Load Balancing algorithms.

11 Recall on Self-Aggregation (1) A Self-Aggregation algorithm is defined as an algorithm capable of enabling a spontaneous formation of groups of compatible nodes. Self-Aggregation

12 Recall on Self-Aggregation () We use the Accurate Self-Aggregation Algorithm previously introduced in Self-Aggregation algorithms for Autonomic Systems (Bionetics 07) It 1 It After more iterations

13 Recall on Dimension Exchange Load Balancing The Dimension Exchange Load Balancing algorithm balances the workload of each couple of nodes until all the nodes have the same load Load-Balancing 1

14 Idea: Combined Approach (1) 1. Group nodes of the same type into homogeneous domains using a self-aggregation algorithm;. Balance the workload among homogeneous domains using the Dimension Exchange Load Balancing algorithm Self-Aggregation Load Balancing 3 9

15 Idea: Combined Approach () Both algorithms can run simultaneously without conflicting; Both algorithms, as other bio-inspired algorithms are able to create the global self-balanced pattern with just those simple local rules Self-Aggregation Load Balancing 3 9

16 Simulation Environment Distributed Simulator written in Java. 0 Monte Carlo Simulations. Network of 100 nodes, average node degree of neighbors. Scale-Free topology. Heterogeneity: 10% (10 types of nodes/jobs). Job distribution: static load of 00 jobs and continuous insertion of 00 jobs every 0s. Node processing time: a) 100% s, b) 70% 7s, 30% 3s (ideal throughput 0jobs/s). Node churn: every 10s 0% of the nodes disappears, and the same number of new nodes appears. We evaluate the throughput and the network load in terms of message overhead.

17 Simulations without Rewiring Number of processed jobs without rewiring Load balancing iterations are inhibited by the fact that the jobs cannot traverse the nodes having different types.

18 Simulations with Rewiring: static load Number of processed jobs with rewiring We have a strong improvement with respect to the previous experiments

19 Simulations with Rewiring: multiple bursts Number of processed jobs with rewiring and multiple bursts The curve grows as a straight line because of the continuous job bursts that are sent to the nodes

20 Simulations with Rewiring: multiple bursts Throughput of processed jobs with rewiring and multiple bursts The throughput is close to the optimal value (the one obtained in homogeneous networks)

21 Simulations: multiple bursts and churn Throughput of processed jobs in presence of a node churn (arrival/departure) of 0%/10secs The throughput is similar to the previous case, therefore it is resistant to this type of uncertainty.

22 Simulations: different service time Throughput of processed jobs in presence of nodes of different processing times Like in the previous case the combined algorithm is still able to obtain nearly-optimal values

23 Simulations: network overhead Message overhead (in terms of number of messages) grows linearly with time in all simulations since it is dominated by the self-aggregation algorithm ( msgs/sec vs msgs/sec).

24 Summary Strenghts: Throughput is close to 0 in all experiments, meaning that the load is spread among almost all workers; Network overhead can be controlled and adapted to the application scenario; System shows self-healing property in presence of churn and dynamism.

25 Summary Strenghts: Throughput is close to 0 in all experiments, meaning that the load is spread among almost all workers; Network overhead can be controlled and adapted to the application scenario; System shows self-healing property in presence of churn and dynamism. A possible problem: Churn may remove key nodes from the network (nodes with high degree) and split the network; However churn and dynamism, especially arrival of new nodes makes this problem transitory.

26 Related Work Related Work on Self-Aggregation: Nodes organize spontaneously into island of similar nodes exploiting their local iterations. Other Self-Aggregation Approaches: topology control in wireless multi-hop networks (Blough et al, 003); overlay organization in Cyclon (Voulgaris et al, 00); evolutive methods based on genetic algorithms (Armor at al, 00). Related Work on Load Balancing: Classical techniques include the work from Cybenko (1989) and its recent extensions (Bahi et al, 007). Another approach modifies the routing infrastructure of a wireless ad hoc network (Nehra et al, 007). However in all these approaches the nodes are assumed to be homogeneous (non-specialized workers).

27 Conclusions We have shown how it is possible to add self-aggregation to the dimension exchange algorithm. This allows to balance the load in networks of heterogeneous nodes, where jobs cannot be processed indifferently by any node and communication between nodes is limited to their logical links. Future Work: Generalize the algorithm to consider nodes that belong to more than a single type at the same time. Study an efficient mechanism to dynamically adapt the frequency of load balancing and clustering iterations. Consider the case in which the environment sends jobs of a given type to nodes of different type.

28 Questions?