Tomography is a computationally intensive process by which the
three-dimensional structure of an object can be reconstructed from a series of
two-dimensional projections. In this thesis, we address on-line execution
of tomography to provide real-time feedback to users collecting data from an
on-line instrument. Context for this work is provided by a powerful electron
microscope located at the National Center for Microscopy and Imaging Research
(NCMIR). Acquiring data from NCMIR's microscope is a lengthy process and is
susceptible to configuration errors. Thus, real-time tomography feedback will
allow users to quickly NCMIR's microscope is a lengthy process and is
susceptible to configuration errors. Thus, real-time tomography feedback will
allow users to quickly identify configuration problems and interact with the
microscope in order to more efficiently acquire data from it. We present an
implementation of on-line parallel tomography which allows for production runs
in Computational Grid environments. Developing applications that leverage
this type of platform is difficult because resources are heterogenous and
dynamic. In our approach, on-line parallel tomography is designed to be
tunable such that it can be configured to adapt to different resource
availabilities. It is coupled with an user-directed, application-level
scheduler which exploits the tunability of the application to determine a
schedule for soft real-time execution. The scheduler utilizes user
constraints, an application model, and dynamic resource load predictions to
determine feasible run-time configurations. The configurations are displayed
as choices to the user where each configuration involves trade-offs between
resolution of the reconstruction, frequency of feedback, and cost of execution.
Once an appropriate configuration is chosen by the user, the scheduler selects
resources, allocates work, and executes the application.
Pre-2018 CSE ID: CS2001-0675
