BALANCE: Signature-BAsed LoAd MaNagement for Loosely Coupled Heterogeneous DistributEd Systems

Cloud Computing;Heterogeneous Distributed Systems;Signature-Driven Load Management;Multi-dimensional Load Pattern Matching

Ramaswamy, Prakash ; Dr. Xiaosong Ma, Committee Chair,Dr. Xiaohui (Helen) Gu, Committee Co-Chair,Dr. Khaled Harfoush, Committee Member,Ramaswamy, Prakash ; Dr. Xiaosong Ma ; Committee Chair ; Dr. Xiaohui (Helen) Gu ; Committee Co-Chair ; Dr. Khaled Harfoush ; Committee Member

University:North Carolina State University

关键词: Cloud Computing;    Heterogeneous Distributed Systems;    Signature-Driven Load Management;    Multi-dimensional Load Pattern Matching;   

Others  :  https://repository.lib.ncsu.edu/bitstream/handle/1840.16/1079/etd.pdf?sequence=1&isAllowed=y

美国|英语

来源: null

【 摘 要 】

Most distributed systems are characterized by modular architecture, distribution of distinct resource (hardware characteristics) and application (workload characteristics) features supporting heterogeneous processing elements. Due to the inherent heterogeneity and distribution nature, entities in distributed systems often tend to have uneven load distribution. In this light, it is imperative to have an efficient load management scheme for heterogeneous distributed systems. Most existing load balancing algorithms determine the appropriate machine for task or process execution using coarse-grained information such as average load on each host. Nevertheless, since distributed systems are characterized by multiple metrics, load balancing algorithms need to consider multi-dimensional resource requirements. More importantly, rather than characterizing system metric using coarse-grained information, the load balancing algorithm needs to consider fine-grained measurements in order to achieve efficient load management for dynamic distributed systems. In this thesis, we present BALANCE, a signature-based load management system to improve resource utilization in dynamic heterogeneous distributed systems. BALANCE dynamically captures fine-grained signatures of dynamic application workloads using time series patterns, performs precise resource tracking and allocation based on the extractedsignatures. BALANCE employs multi-dimensional time series indexing and uses Pastry, an existing scalable peer-to-peer scalable storage system to achieve efficiency and scalability respectively. We implement a prototype of BALANCE and deploy it on the PlanetLab and NCSU Virtual Computing Lab (VCL). Our experiments show that BALANCE completesefficient task allocation and hence load balancing within tens of milliseconds. It improves the overall request satisfaction rate by 30-80% compared to the existing approaches.