the efficient characterization of adaptive parallel applications is usually challenging due to their complexity and large problem size. Unlike traditional profiling approaches which target the tracing of events or det...
详细信息
ISBN:
(纸本)9780889866386
the efficient characterization of adaptive parallel applications is usually challenging due to their complexity and large problem size. Unlike traditional profiling approaches which target the tracing of events or determining performance parameters for subroutines, the approach described in this paper attempts to discover the inherent adaptivity of parallel applications mapped to the computation domain/mesh, which are independent of runtime environment, so as to aid in the performance tuning of parallel applications, especially dynamic load balancing and repartitioning. Our profiling scheme only requires one-time execution of the target program on any platform to generate a sequence of traces with timestamps. the traces can then be fed to simulations under various system configurations while independent of the real application. Preliminary experiments have been performed to evaluate the proposed profiling techniques.
the Replica Placement Problem (RPP) aims at selecting the nodes for duplicating data objects in order to optimize their access. Even though a lot of work already exists on RPP, the issue of implementing the resulting ...
详细信息
ISBN:
(纸本)9780889866386
the Replica Placement Problem (RPP) aims at selecting the nodes for duplicating data objects in order to optimize their access. Even though a lot of work already exists on RPP, the issue of implementing the resulting allocation scheme is typically overlooked. In this paper we introduce the Replica Transfer Scheduling Problem (RTSP), briefly stated as: given a network of servers with limited storage capacitv, a set of data objects and two replication schemes x(old) and X-new, find a schedule of object transfers and deletions for implementing X-old based on X-old with minimum communication cost. Given that this problem is NP-complete, we introduce several different heuristics to solve it, and evaluate them via simulations.
this work concerns the use of machine learning techniques (genetic algorithms) to optimize load balancing policies in the openMosix distributed operating system. Parameters/alternative algorithms in the openMosix kern...
详细信息
Constraint-based synchronization pioneered by (concurrent) logic and concurrent constraint programming is a powerful mechanism for synchronizing concurrent computations. this paper describes (1) a general constraint-b...
详细信息
Complex system design and analysis requires simulations with different views or domains. Multi-domain simulation models typically are a collection of simulation model integrated through various means. these concurrent...
详细信息
ISBN:
(纸本)9780889866386
Complex system design and analysis requires simulations with different views or domains. Multi-domain simulation models typically are a collection of simulation model integrated through various means. these concurrent models may be implemented to be closely integrated or distributed as parallel-distributed simulation or mix of the two. Performance of a coupled simulation between multiple domain models depends on various factors. Some of them are, integration backplane architecture, messaging schemes, sharing scheme, level of coupling, fidelity of domain models etc. Framework for Optimizing Multidomain Simulation (MSOMS) is a framework which can be used to estimate performance and effectiveness of integration methods under various scenarios. the framework is referred as meta-simulation, since it is a "simulation of a simulation". this paper captures formulation and demonstration of MSOMS. We compared two possible implementations of a multi-domain model using MSOMS as a case study.
In a grid computing environment, dynamicity, and geographically distributed sites, make task scheduling problems challenging to solve. It is hard for a local site to obtain precise real-time information about other si...
详细信息
ISBN:
(纸本)9780889866386
In a grid computing environment, dynamicity, and geographically distributed sites, make task scheduling problems challenging to solve. It is hard for a local site to obtain precise real-time information about other sites given that specific information on a site such as load and computing resources may change rapidly. Moreover, in data grid environment, large scale data intensive applications make scheduling problems even more challenging since both computational and data storage resources must be taken into consideration. In this paper we propose an innovative peer-to-peer scheduler to solve these problems. this scheduler is distributed and scalable. We used simulation to evaluate the performance of the scheduler under different circumstances, such as different number of hops to search suitable sites and different number of incoming tasks. Results show that our scheduler can successfully schedule around 75% of incoming tasks within their deadlines in average. For computation-intensive tasks, it can successfully schedule more than 90% of incoming tasks.
A future is a parallel programming language construct that enables programmers to specify potentially asynchronous computations. We present and empirically evaluate a novel implementation of futures for Java. Our futu...
详细信息
ISBN:
(纸本)9780889866386
A future is a parallel programming language construct that enables programmers to specify potentially asynchronous computations. We present and empirically evaluate a novel implementation of futures for Java. Our futures implementation is a JVM extension that couples estimates of future computational granularity with underlying resource availability to enable automatic and adaptive decisions of when to spawn futures in parallel or to execute them sequentially. Our system builds from, combines. and extends (i) lazy task creation and (ii) a JVM sampling infrastructure previously used solely for dynamic and adaptive compilation. We empirically evaluate our system using different benchmarks, triggers for automatic spawning of futures, processor availability, and JVM configurations. We show that our future implementation for Java is efficient and scalable for fine-grained Java futures without requiring programmer intervention.
An efficient range query processing support is required for distributed Hash Table (DHT)-based P2P networks as consistent hashing destroys the inherent order of numeric keys. In this paper, we present a lightweight an...
详细信息
ISBN:
(纸本)9780889866386
An efficient range query processing support is required for distributed Hash Table (DHT)-based P2P networks as consistent hashing destroys the inherent order of numeric keys. In this paper, we present a lightweight and efficient mechanism, called Range Hash Tree (RHT), to support range queries on DHTs. In RHT, key space is partitioned into small ranges such that data items within one range can be stored on one node. To quickly resolve a range query, we introduce a scalable encoding algorithm that can represent the key space partitioning status with a small RHT Signature. Compared to other approaches, RHT provides a bounded query delay, which is independent of the size of range query and the number of matching data items. Our experiments show that RHT works efficiently on both uniform and much skewed data distributions.
Space filling curves (SFCs) are proximity preserving linearizations of multidimensional data. they are frequently used for parallel domain decomposition in scientific computing applications where interactions occur be...
详细信息
A two-step Modified Critical Path and Area-based (MCPA) scheduling heuristic is developed which targets at improving the processor allocation phase of an existing Critical Path and Area-based (CPA) scheduling algorith...
详细信息
暂无评论