Parallel programming in the cloud is a research challenge that has been
intensively studied, motivated mostly by the potential impact that it can have
on the performance of scientific computations in areas such as Chemistry
(Molecular Dynamics), Biology (Genetics) and O&G, among others. Despite the
various research initiatives, efficient solutions for parallel programming in
the cloud still require a complex combination of programming models and languages,
which often do not produce the desired level of performance. The authors of
this work have proposed OmpCloud (ompcloud.org), a program- ming model that
extends OpenMP to easy the task of writing parallel pro- grams to the cloud.
OmpCloud produced good performance numbers when running applications in the
clouds of Amazon AWS and Microsoft Azure. OmpCloud was also extended to enable
the execution of FPGA bitstreams on Intel HARP and Xilinx AWS-F1 architectures
(hardcloud.org). This talk defends the idea that a task parallelism extension of
OmpCloud has the potential to produce an efficient and seamless cloud programming
model, when supported by efficient data distribution and scheduling mechanisms.
This could be achieved by pursuing two major goals. First, by leveraging on
OpenMP task parallelism to extend the current OmpCloud model. Second, by
designing a runtime which could encapsulate efficient MPI based data distribution,
fault tolerance and load balancing. This approach could result in a simpler and
more efficient cloud programming paradigm, that has the potential to easy the
task of programming complex scientific applications. We intend to evaluate this
approach using a number of scientific programs from Molecular Dynamics and O&G
benchmarks.