Research
Structure Formation
Clusters
Subgrid Models
Resources
Codes
As a bare minimum, an astrophysical code needs to be able to evolve collisionless particles in a self gravitational field, i.e. to solve Vlasov-Poisson equation for a given system. However, that usually represents only the first step when one is concerned about accuracy. To better represent physical world, simulations are including more and more of nongravitational physics, like hydrodynamics, magnetic fields, heating/cooling, radiation... All these units need to be efficiently written, thoroughly tested, documented, and regularly maintained. Thus, today's general purpose codes are developed and maintained by whole groups, rather than just one individual. One such example is a FLASH code. But still, it is usually very helpful to have a small code, meant to attack one particular problem where it achieves high efficiency, reliability, and robustness. As a bonus such codes are easy to understand and change. For large-scale structure studies we have MC2, as such code.
FLASH
Type: All purpose mesh code
Geometry:
Cartesian 1D,2D,3D; cylindrical 2D, 3D; spherical
1D, 2D, 3D; polar 1D, 2D
adaptive mesh
Physics:
gravity, hydrodynamics, relativistic hydrodynamics,
magnetohydrodynamics, nuclear reactions, radiation, conductivity, viscosity.
Developed and maintained at University of Chicago
MC2
Type: Particle-mesh code for cosmological structure formation
Geometry:
Cartesian 3D
fixed mesh
Physics:
gravity, simplified hydrodynamics (HPM).
Developed and maintained at Los Alamos National Laboratory
Computers
Computational requirements from simulations are extremely high in these days, and they
are going to be even higher as more spatial and mass resolution is needed to
accurately capture a phenomenon under consideration. Besides, to obtain better results,
simulations are becoming more realistic, including more physics.
The good thing is that computational power is steadily increasing for few decades already
(Moore's law), but still,
even the best personal computers today are not matching scientific needs. They are good
for code development, but serious runs require serious power. To take the extreme example,
recently done Millennium simulation
for its calculation had required 350 000
processor hours (one month on 512 processors), 1 TB of RAM memory, and it produced 23 TB of output
data. This is clearly a great technical challenge, and besides highly optimized code, it
needs state-of-the-art supercomputers to be performed.
Some of computers I am using for the research are presented below.
Tungsten
Architecture: Linux Cluster
Check status and jobs:
Cluster Monitor
Hardware:
Dell PowerEdge 1750 server
1450 nodes - 2900 Xeon 3.2 Ghz processors
3 GB RAM per node
70 GB hard disk per node + 122TB shared
Myrinet 2000 ethernet
Software:
Linux Red Hat 9.0
LSF job manager
MPI, Intel Fortran 77/90/95 C C++
Peak performance (Linpack benchmark): 9.8 teraflops - 4th
fastest in the world at the release (2003).
Location: National Center for Supercomputing Applications, Urbana, IL
Cobalt
Architecture: Shared Memory System
Check status and jobs:
Cluster Monitor
Hardware:
SGI Altix system
1024 Intel Itanium 2 processors
3 TB global RAM
370 TB storage
Software:
SGI ProPack 3.4 (kernel 2.4)
PBS Pro job manager
MPI, Intel Fortran 77/90/95 C C++
Peak performance (Linpack benchmark): 6.1 teraflops - 48th
fastest in the world at the release (2005).
Location: National Center for Supercomputing Applications, Urbana, IL
ASCI QSC
Architecture: UNIX Cluster
Check status and jobs: N/A
Hardware:
HP/Compaq Alphaservers ES45
256 nodes - 1024 processors
16 GB RAM per node
15 TB storage
Quadrics Interconnect network
Software:
Tru64 UNIX
PBS job manager
MPI, Compaq Fortran 90/95/HPF C C++
Peak performance (Linpack benchmark): 2.56 teraflops;
the whole Q: 13.8 teraflops - 2nd
fastest in the world at the release (2003).
Location: Los Alamos National Laboratory, NM
|
My Universe > Science > Resources |
© Zarija Lukic 2006 |