|
|
|
|
|
High Level Programming for Computational Grids
- This project is a 2-year British Council/DAAD funded travel grant
(Project No. 1223) partners at LMU Munich,
Phillips-Universitaet Marburg, and St Andrews University. We aim to
evaluate a single large program on a computational grid, i.e. on a
collection of grid- enabled workstation clusters. This entails
developing a sophisticated language implementation that adapts
dynamically to such a hierarchical, heterogeneous and high-latency
architecture.
- (06/2003 - 06/2005)
- Special purpose High Performance Computers (HPCs) are
expensive and rare, but workstation clusters are cheap and
becoming common. Emerging GRID technology offers the
opportunity to integrate GRID-enabled clusters into a single
HPC. The acceptance of GRIDs, however, is seriously hampered
by the lack of security guarantees and the difficulty of
efficiently managing the parallelism in such a heterogeneous
clusters, with characteristics radically different from a
conventional HPC: it is shared with components used
simultaneously by multiple users; it is heterogeneous
connecting clusters of different sizes and speeds; it is
hierarchical with communication much faster within the
cluster than between clusters; it has high communication
costs(latency); the effective speeds of components may vary
during program execution, due to sharing nodes and
connections with other users.
- To program this complex and dynamic architecture
effectively we propose to use a language with high-level
constructs supported by a sophisticated runtime environment
(RTE) that automatically adapts to the dynamically-changing
execution environment. To meet basic security concerns in
an open GRID environment we propose to use novel analysis
techniques to provide information about resource
consumption. Key components of the environment are a
monitoring tool that determines static and dynamic
properties of the network of HPCs, a distributed
infrastructure for attaching resource certificates to
executable code and a virtual machine that dynamically
adapts parallel execution based on the current
configuration and the provided resource certificates to
improve performance.
- The effectiveness of the
adaptive RTE will be evaluated by measuring the performance
of two existing substantial parallel applications on
several networks of GRID- enabled clusters. The monitoring
tool, certification system and adaptive virtual machine
have separate research objectives and will be developed
separately before being integrated for evaluation.
- We aim to determine whether a GRID of cheap clusters can be
transparently and effectively utilised as a single high performance
computing platform using a high-level language supported by a
sophisticated runtime environment (RTE). The core objective is
determining whether a sophisticated RTE can be constructed to deliver
speedups and scaleups for real programs on modest GRIDs of high
performance computers (HPCs). The RTE comprises a monitoring tool and
an adaptive virtual machine, supported by resource certificates that
are attached to executable code when sent to another processor. The
main research objectives are as follows.
- Monitoring tool. To determine the crucial static and dynamic
architectural properties of a GRID-of-HPCs at the correct level of
abstraction to facilitate effective management of high performance
computing. To construct and evaluate prototype tools using the GRID
infrastructure.
- Adaptive virtual machine. To port an existing adaptive RTE
to the GRID and radically extend it with new strategies to
automatically and dynamically manage program execution on a
GRID-of-HPCs. The management strategies utilise static and dynamic
configuration information produced by the monitoring tool, and
guaranteed resource use information. The strategies must effectively
manage the hierarchical, heterogeneous, shared and high-latency
nature of the architecture.
|
|
