A probabilistic framework for estimation of execution time in heterogeneous computing systems
Abstract
A heterogeneous computing (HC) system provides a variety of architectural capabilities, orchestrated to perform an application whose subtasks have diverse execution requirements. An HC system may include a suite of independent machines of different types interconnected by a high-speed network. Static matching (assigning subtasks to machines) and scheduling (ordering the execution of subtasks assigned to the same machine) are more important for HC systems than for distributed homogeneous systems, because dynamic subtask migration is generally difficult and impractical in HC systems. A key component in static matching and scheduling research is predicting the performance associated with a given matching/scheduling choice. Most of the past approaches for performance prediction, in this context, are based on deterministic models. In this thesis, a probabilistic framework is developed, which offers a more realistic characterization of performance than deterministic models. A methodology is first introduced for estimating the execution time distribution for a given data parallel program that is to be executed in either the synchronous SIMD mode, the asynchronous SPMD mode, or an SIMD/SPMD mixed-mode environment. This approach serves as the basis for modeling the execution time distribution for an entire task graph to be executed in a general HC environment. In this model, the task graph consists of data parallel subtasks; sequential subtasks are conveniently modeled by assuming a single-processor SPMD platform. It is shown that, for a given matching and scheduling, computing the exact distribution of the overall execution time of a task graph is very time consuming, and thus impractical. A proposed approach is provided that approximates this distribution and requires a relatively small amount of calculation time. An upper bound for the difference between the exact distribution and that provided by the proposed approach is derived. From this bound it is shown that the proposed approach approximates the exact distribution with high accuracy in many practical situations; conditions are also determined for which the distribution of the proposed approach equals the exact distribution. Extensive numerical studies are performed to further validate the utility of the proposed methodology.
Degree
Ph.D.
Advisors
Antonio, Purdue University.
Subject Area
Electrical engineering|Computer science|Systems design
Off-Campus Purdue Users:
To access this dissertation, please log in to our
proxy server.