Load Balancing in a Cluster-based Active Jiani Guo (Student Member, IEEE) Laxmi Bhuyan (Fellow, IEEE) March 15 th 2005 Seo, Dong Mahn
March 15th 20052/35 Contents I. Introduction II. Background and Related Work - Background, Related work III. Cluster-Based Active Router Architecture - Cluster-Based Active Router, Cluster-based Multimedia Transcoding Service, Software Framework of Active Router Cluster, Load Test Mechanism IV. Load balancing Techniques - First Fit (FF), Stream-based Mapping (SM), Least Load First (LLF), Adaptive Load Sharing (ALS) Policy, Adaptive Portioning (AP) V. Design of the Experiments - Experimental Settings, Design Issues VI. Performance Metrics - System Scalability, Load Sharing Overhead, Video Quality VII. Performance Evaluation - Effect of Varying Load test Epoch, Effect of Varying Packet Size, Effect of Varying Number of Dispatchers, System Throughput, Video Quality VIII. Conclusion
March 15th 20053/35 Introduction (1) Parallel and distributed computing systems Active network architecture NetScript project, ANTS system, SwitchWare Mega project, Journey network Multimedia streaming To provide real-time transcoding service Cluster-based Active Router Architecture (CLARA)
March 15th 20054/35 Introduction (2) Load balancing Simple static policy Adaptive load balancing policy Same flow packets Multimedia Active Router reduce the transcoding time ensure that the out of order departure of packets belonging to ther same stream Round-robin & one flow/stream Inter-departure time & Jitter
March 15th 20055/35 Introduction (3) In this paper Linux-based active router cluster Gigabit Ethernet, multithreaded software architecture Manager node, Computing Servers Load test mechanism Adoptive load sharing technique based on the HRW algorithm New algorithm : Adaptive Partitioning (AP) First fit, Adaptive load sharing schime, Stream-based mapping, AP
March 15th 20056/35 Background Transmission of multimedia information To store multiple copies of the source stream JPEG transcoding technique, determinig the level of transcoding Cluster-based web distillation proxy (TranSend) Layered source-coding algorithm Real-time transcoding service Application level video gateway architecture, active network node (ANN), proxy architecture, CLARA
March 15th 20057/35 Related work Load balancing Simple static policy Random distribution policy, modulus-based round robin policy Load balancing algorithms with the concept of flows HRW algorithm MPEG transcoding Load sharing schemes Round robin, stream-based round robin, adaptive load sharing
March 15th 20058/35 Cluster-Based Active Router
March 15th 20059/35 Cluster-based Multimedia Transcoding Service MPEG-1, MPEG-2, H.261/263, MPEG-4 MPEG-4 Object Composition Petri Net (MOCPN) To deploy transcoding among different coding schemes Ours To provide a general transcoding To develop a scheduling technique Fast transcoding Reducing the out-of-order departure
March 15th /35 Software Framework of Active Router Cluster (1)
March 15th /35 Software Framework of Active Router Cluster (2)
March 15th /35 Load Test Mechanism
March 15th /35 First Fit (FF) Round robin way Scheduler blocked, if full Put the media unit, if vacancy Without extra load analyzer Better than the simple round-robin scheme
March 15th /35 Stream-based Mapping (SM) To preserve the computation order To keep the simplicity of first fit Propose and implement f(c) = c mod N, c is the stream number, N is the total number of servers. j = i % N, i is stream, j is dispatch queue. Good for homogeneous servers and some specific input patterns M N, M = multiple of N (M is the number of streams, N is the number of servers)
March 15th /35 Least Load First (LLF) Actual load condition Picks the currently least loaded server A i (t), the number of outstanding requests Transcoding workload is proportional. The same stream are possible to be distributed to different servers High jitters at destination
March 15th /35 Adaptive Load Sharing (ALS) Policy Extended HRW technique v is the identifier vector of the packet, j is the server node to which the packet Utilization of each server Implement function Same weights load information load statistics information modify and recalculate
March 15th /35 Adaptive Portioning (AP) (1)
March 15th /35 Adaptive Portioning (AP) (2)
March 15th /35 Adaptive Portioning (AP) (3)
March 15th /35 Experimental Settings Assumptions 1 input port, 1 output port Multiple media streams concurrently on the Manager node Departure-Recorder program MPEG-1 color video to black/white, FFMpeg No multi-layer encoding, stream error correction encoding 15 frames in a GOP, 30fps, around 50KB of GOP, a GOP per 0.2 sec
March 15th /35 Design Issues 1) Load Test Epoch Sensitive parameter for feedback-based load balancing schemes 2) Number of Threads Thread switching is non-negligible overhead. 3) Packet Size 10K to 100K GOP size Packet size
March 15th /35 System Scalability Measure the scalability of our cluster System throughput, GOPs/sec One of the most important metrics
March 15th /35 Load Sharing Overhead Metric 1 : Load Test Overhead is the average time consumed by the Manager node to poll through all servers to collect the load statistics information. Metric 2 : Load Remapping Overhead is the time consumed to set the current load for each server in a specific load balancing scheme.
March 15th /35 Video Quality Metric 1 : Departure Jitter per Stream is the standard deviation of the interdepature time among GOPs when the stream is received by the Departure-Recoder PC. Metric 2 : Average Interdeparture Time among GOPs per Stream is the mean of the interdepature time among GIOs when the stream is delivered to the active router. Metric 3 : Out-of-Order Rate per Stream describes how many GOPs among all the GOPs in a stream depart out of order.
March 15th /35 Effect of Varying Load test Epoch
March 15th /35 Effect of Varying Packet Size (1)
March 15th /35 Effect of Varying Packet Size (2)
March 15th /35 Effect of Varying Number of Dispatchers
March 15th /35 System Throughput
March 15th /35 Video Quality (1)
March 15th /35 Video Quality (2)
March 15th /35 Video Quality (3)
March 15th /35 Conclusion (1) Active network New direction for processing network traffic Distribute expensive computation Multimedia traffic Cluster-based active router architecture Design, implement, evaluate Load sharing policy First Fit, Least Load First, Stream-based Mapping, Adaptive Load Sharing, Adaptive Partioning
March 15th /35 Conclusion (2) Load balancing scheme To improve the throughput First Fit and Least Load First schemes To achieve good video playout quality Stream-based Mapping, Adaptive Load Sharing and Adaptive Partitioning Feedback-based scheme Adaptive partitioning algorithm Future research Address the problem Explore how to guarantee QoS requirements for different streams
March 15th /35 Q & A