Simulation study of a parallel processor with unbalanced loads. Master's thesis
Technical Report
·
OSTI ID:5180989
The purpose of this thesis was twofold: to estimate the impact of unbalanced computational loads on a parallel-processing architecture via Monte Carlo simulation; and second to investigate the impact of representing the dynamics of the parallel-processing problem via animated simulation. It is constrained to the hypercube architecture in which each node is connected in a predetermined topology and allowed to communicate to other nodes through calls to the operating system. Routing of messages through the network is fixed and specified within the operating system. Message-transmission preempts nodal processing causing internodal communications to complicate the concurrent operation of the network. Two independent variables are defined: 1) the degree of imbalance characterizes the nature or severity of the load imbalance, and 2) the degree of locality characterizes the node loadings with respect to node locations across the cube. A SLAM II simulation model of a generic 16 node hypercube was constructed in which each node processes a predetermined number of computational tasks and, following each task, sends a message to a single randomly chosen receiver node. An experiment was designed in which the independent variables, degree of imbalance and degree of locality were varied across two computation-to-IO ratios to determine their separate and interactive effects on the dependent variable, job speedup. ANOVA and regression techniques were used to estimate the relationship between load imbalance, locality, computation-to-IO ratio, and their interactions to job speedup. Results show that load imbalance severely impacts a parallel-processor's performance.
- Research Organization:
- Air Force Inst. of Tech., Wright-Patterson AFB, OH (USA). School of Engineering
- OSTI ID:
- 5180989
- Report Number(s):
- AD-A-189573/9/XAB; AFIT/GCS/ENG-87D-20
- Country of Publication:
- United States
- Language:
- English
Similar Records
A performance study of the hypercube parallel processor architecture
Development of a low-latency scalar communication routine on message-passing architectures
Message passing kernel for the hypercluster parallel processing test bed
Journal Article
·
Thu Feb 28 23:00:00 EST 1991
· Simulation; (United States)
·
OSTI ID:6098166
Development of a low-latency scalar communication routine on message-passing architectures
Thesis/Dissertation
·
Mon Jan 10 23:00:00 EST 1994
·
OSTI ID:10126154
Message passing kernel for the hypercluster parallel processing test bed
Conference
·
Sat Dec 31 23:00:00 EST 1988
·
OSTI ID:5805456