Measuring Network Performance of Multi-Core Multi-Cluster (MCMCA) Norhazlina Hamid nh3g11@ecs.soton.ac.uk Supervisor: R J Walters and G B Wills PUBLIC SERVICE DEPARTMENT OF MALAYSIA Outline Introduction and Definition Motivation Related Work Research Objectives The Architecture Methodology Conclusion and Progress Work 2 Introduction The emergence of High Performance computing (HPC) that includes Cluster computing has improved the availability of powerful computers and high speed network technologies. The implementation of multi-core cluster as a platform of high performance network will supports high availability and enables scalability of the networks. 3 Definition Multi-Core Multi-Cluster Architecture (MCMCA) is a collection of multiple multicore cluster, interconnect by a network. MCMCA are built from: Cluster computing is a collection of computer, interconnect by a network which work together to form a single computer 4 Definition A multi-core cluster is a cluster where some or all the nodes in the cluster have multicore processors Multi-cluster architecture is a multiple cluster system that is connected via the cluster interconnection networks Cluster computi ng Multicore process or Multicluster MCMCA 5 Motivation Existing models do not address the potential performance of the interconnection networks within large-scale multi-core clusters . A high communication latency of interconnection network can dramatically reduce the efficiency of the cluster system. Scaling up by adding more processors to each cluster increase the processing power.6 Related Works Furhad et al. (2013) proposed the EnMesh topology to address the issue of communication delays that occur with long distance communication within the remote nodes of a NoC. Mohsen et al. (2013) proposed a new mapping technique to assign parallel processes into processing cores of a multicore cluster which based on queuing network modelling of a limited-size cluster. 7 Objectives To propose a new architecture for multicore multi-cluster. To investigate the performance of interconnection networks of multi-core multi-cluster architecture. To demonstrate the feasibility of the proposed architecture by computer simulation experiment and measurements. 8 The Architecture of MCMCA Extended Multi-core Cluster System Architecture (EMCSA) Cluster 0 Inter-Chip Intra-Chip Core Core Core Core Cluster 1 Inter-Chip Intra-Chip Core Core Core Intra-Chip Core Inter-Chip Core Core Core Core Inter-Chip Intra-Chip Core Core Core Core Cache Cache Cache Cache Cache Cache Cache Cache Cache Cache Cache Cache Cache Cache Cache Cache Dual-Core Chip Dual-Core Chip Dual-Core Chip Dual-Core Chip Dual-Core Chip Dual-Core Chip Dual-Core Chip Dual-Core Chip Memory Memory Memory Memory Memory Memory Memory Memory Key: Intra-Chip : Inter-Chip : Intra-Cluster Network : Inter-Cluster Network : Multi-Cluster Network: Intra-Cluster Network (ACN) Intra-Cluster Network (ACN) Inter-Cluster Network (ECN) Inter-Cluster Network (ECN) Multi-cluster Network (MCN) 9 Communication Network There are five communication networks in MCMCA 1. 2. 3. 4. 5. IntrA-chip Communication network (AC) IntEr-chip Communication network (EC) IntrA-Cluster Network (ACN) IntEr-Cluster Network (ECN) Multi-Cluster Network (MCN) 10 IntrA-chip Communication network (AC) Message passing between two processor cores on the same chip 11 IntEr-chip Communication network (EC) Message passing across processors on different chips but within a node 12 IntrA-Cluster Network (ACN) Message passing between processors on different nodes but within the same cluster 13 Communication Network Message passing between clusters 14 Methodology Computer simulation experiments using OMNeT network simulation tool. Focus on network latency to measure the performance. Early stage experiments are based on a single-core multi-cluster architecture. The experiments have been performed using given configuration and parameters of previous paper to create baseline results.15 Experiment Results Results from the experiments have shown that as the traffic rate increases, the average communication latency increases following the assumptions that the messages are delayed by having to wait for resources before traversing into a network. The results confirm that the simulation model is a good basis to measure the communication latency for a large-scale cluster, and can be extended to multi-core 16 multi-cluster architecture. Methodology Network Latency (∑NL) are obtained from: – Total Network Latency of Message Passing Within Cluster (L) – Total Network Latency of Message Passing Between Clusters (L) – Probability of outgoing message in a cluster (P) – Probability of message in a cluster (1-P) Analytical Formula: ∑NL = (L x (1-P)) + (L x P) 17 Methodology Both Network Latency (L and L) are obtained from: – Waiting time at the source node (W) – Delay while traversing the network(D) – Time for each packet to reach its destination node (E) – Waiting time at transfer switch between clusters (Wts) Analytical Formula: L = W + D + E L = W + D + E + 2Wts 18 Conclusion Presenting a new architecture for measuring the performance of interconnection networks in MCMCA. Progress work: 1. Developing a MCMCA simulation model. 2. Conduct a simulation experiments with different network technology and bandwidth. 3. Developing analytical equation for the architecture. 19 Published Paper 1. Hamid, Norhazlina, Walters, Robert John and Wills, Gary Brian (2014) Performance evaluation of multi-core multi-cluster architecture. In, Emerging Software as a Service and Analytics, Barcelona, ES,03 - 05 Apr 2014. Scitepress9pp, 4654. 2. Hamid, Norhazlina, Walters, Robert John and Wills, Gary Brian, " An Architecture for Measuring Network Performance in MultiCore Multi-Cluster Architecture (MCMCA). In, Euro-Asia Conference on Computational Intelligence and Communication Networks, Antalya, Turkey, 21-23 Apr 2014." 3. Hamid, Norhazlina, Walters, Robert John and Wills, Gary Brian, " An Architecture for Measuring Network Performance in MultiCore Multi-Cluster Architecture (MCMCA)," International Journal of Computer Theory and Engineering vol. 7, no. 1, pp. 57-61, 20 February 2015. Q&A Thank you. 21