Preface
With evolution of mathematical modeling and creation of high-performance
computer systems, many scientific applications have appeared that demand
increasing computational performance, higher than any of available supercomputers
can provide. In particular, for the super large scale numerical modeling
it is necessary to integrate several supercomputers, i.e., to create a
Grid. Not any application can be well solved on grids because of slow
communications. However, such application problems as search for alien
civilizations, prime numbers search and climate prediction are successfully
running on grids.
Another problem is a rapid progress in microprocessor development that
forces us to use heterogeneous computer systems for solution of the large-scale
problems. In particular, in 2007 the Siberian Supercomputing center (Novosibirsk,
Russia) exploits the following multicomputers: the 32 processors MVS-1000,
based on the alpha microprocessor (833 Mhz), the 128 processors MVS-1000,
based on the alpha microprocessor (633 Mhz) and the 60 processors HP cluster,
based on the Intel Itanium II microprocessor. There are also two clusters
based on Intel Pentium III and Opteron in Novosibirsk State University
that are used in scientific calculations. Therefore, there is a necessity
to create the software that will provide the large-scale simulation in
heterogeneous environments. For now, numerous GRID projects oriented to
different applications are under development.
The speed of communication is permanently growing and now it is possible
to organize numerical simulation on GRID of multicomputers. In 2004, NumGRID
project intended for the creation of the necessary grid system software
started in Novosibirsk (ICM&MG).
Objectives
The main objective of the NumGRID project is to provide the use of remote
multicomputers for large-scale numerical simulations
Several large-scale numerical applications were already solved in NumGRID
environment:
1. The simulation of protoplanetary disc evolution and galaxy formation.
2. Digital Electromagnetic Model of the Power System: Parallel Implementation
for Multicomputers.
3. A numerical model for shallow-water flows: dynamics of the eddy shedding,
WSEAS Transactions on Environment and Development.
NumGRID middleware
The NumGRID middleware is a collection of tools that helps to join several
computational clusters, which can be administered independently, to create
more powerful computational resource. The NumGRID provides users with
a partial implementation of MPI capable to spread MPI-applications over
nodes of independent clusters, computational resource and job management
system with convenient user interface and a library that allows for easy
development of parallel programs with dynamic properties
The NumGRID middleware consists of 5 components:
1. NumGRID_MPI library (partial MPI support for Grid, including
support for MPICH, LAM-MPI, ...)
2. NumGRID_jobmanager (application management, MPI message routing,
queue systems support - SGE, PBS, ...)
3. Debugging and monitoring subsystem
4. Cross-platform user interface (client, server, security subsystem
based on SSL)
5. libAPT (support for development of numerical applications: a
growing collection of data structures specific for numerical simulations
and supporting computational and communicational patters specific for
this area of applications)
NumGRID features
1. Multicomputers/clusters are included into NumGRID
2. Each node of a multicomputer can be an SMP system (2 processors or
more)
3. MPI programs can be executed on NumGRID without any changes. Global
addressing of all the NumGRID resources is provided.
4. Automatic providing of the dynamic properties of application programs
(tunability, dynamic load balancing, program execution monitoring, reliability)
5. Security and safety of calculations on the Grid are provided based
on SSL
6. For any application a unique set of multicomputers can be linked in.
Also NumGRID environment allows all multicomputers to keep their administrative
policies unchanged
Support for development of numerical applications
- The idea:
In numerical simulations, the parallel solution to a problem usually
comes from data decomposition. Thus, the idea in the base of the libAPT
library is to collect distributed implementations of data structures
typical for numerical simulations.
- The principles:
- Data structures must be distributable and provide users (application
programmers) with different levels of control ranging from automatic
execution to detailed planning of distribution by the user.
- The control of distribution should be expressed by the user in the
high level terms appropriate for the data structures. Thus, working
with multidimensional array, the appropriate terms are: an array, a
layer, an element, a column, a block, not bytes, doubles, memory locations,
pointers, etc.
- Dynamic redistribution of data structures must be supported with different
levels of control as well.
- Data structures that are used together in numerical simulation should
be implemented with possible combined use in mind.
- Currently implemented:
- Distributed multidimensional data array that can be specialized by
any type of element. Several data arrays in a program can be bound to
each other and be distributed appropriately. Dynamic redistribution
of the arrays is supported. The user can control distribution of the
data structures at different levels up to providing his own algorithms
for redistribution planning. Reduction operations, input/output, exchange
of the boundary values at cross-sections are supported.
- Particle list that can be bound to a data array and be partitioned
automatically according to the partitioning of data arrays. This data
structure allows for implementation of Particle-in-Cell simulation methods.
Publications
[1] N.V.Malyshkin, B.Roux, D.Fougere, V.E.Malyshkin. The NumGRID metacomputing
system. In Bulletin of the Novosibirsk Computing Center, series Computer
Science, issue 21(2004), pp.57-68
[2] D.Fougere, M.Gorodnichev, N.Malyshkin, V.Malyshkin, B.Roux. NumGRID
software for MPI based applications. // BULLETIN of the Novosibirsk Computing
center series: Computer Science, issue 22(2005), NCC Publisher, Novosibirsk,
2005, pp. 41-51.
[3] D.Fougere, M.Gorodnichev, N.Malyshkin, V.Malyshkin, A.Merkulov, B.Roux.
NumGrid Middleware: MPI Support for Computational Grids. // In Proceedings
of the PaCT-2005, Springer Verlag, LNCS series, Vol. 3606, Krasnoyarsk,
2005, pp. 313-320.
Download
(Currently in russian only, without libAPT)
NumGRID package (v 1.0), 2007
NumGRID userguide (v 1.0), 2007
|