Numerical
Simulation Using Beowulf Systems
Parallel Computers in the CFD Lab at OSU
n
A cluster of PCs interconnected by Local Area
Network (LAN).
n
Running an open-source UNIX-like operating system, e.g. Linux.
n
Executing parallel applications programmed with
an industry standard message passing model and library, e.g. Message Passing
Interface (MPI).
Why Beowulf?
n
Availability:
Constructed by commodity PC parts,
available everywhere.
n
Scalability: To
improve performance, just include additional nodes.
n
Cost-effectiveness: Cost of PC and networking parts continues
lowering.
cese.eng.ohio-state.edu
32 nodes, built in 2003
Hardware:
CPU: Pentium
4 2.8GHz
Memory: 1 GB PC3200 DDR memory
Hard Drive: 73 GB SCSI for master node, 80 GB
IDE for slave nodes
Motherboard:
Network card: onboard Intel 82547EI Ethernet controller
Network Switch:
SMC
8624T 24-port Gigabit switch
KVM Switch: Aten ACS-1216AL
Software:
Operating System: RedHat Linux 9
Compiler: Intel
Fortran Compiler 7.1 for Linux
MPI: Lam/MPI
7.0
Cost: 60K
Previous Systems:
Generation I:
2 nodes, built in 1998
Hardware (Dell Precision workstation):
CPU: dual
Xeon 450 MHz CPU
Memory: 768
MB SDRAM ECC
Network card: 3Com
905 Fast Ethernet card
Software:
Operating System: SuSE Linux 6.1
Compiler: GNU
Fortran compiler (included in all Linux distribution)
MPI: MPICH 1.2
Cost: 24K
Generation II:
8 nodes, built in 1999
Hardware:
CPU: Dual Intel Pentium III 450MHz
Memory: 512 MB SDRAM ECC
Hard drive: 13 GB IDE for master node, 6 GB IDE for slave nodes
Motherboard:
Network card: Dual 3Com 905C Fast Ethernet (100 Mbps) with channel bonding technology
Network Switch: Dual SMC-EZ1024FDT 16-port 10/100 Mbps
KVM Switch: Iogear GCS138
Software:
Operating System: RedHat Linux 7.1
Compiler: PGI Cluster Development Kit (CDK) 3.0
MPI: MPICH 1.2
Cost: 25K
Generation III:
8 nodes, built in 2002
|
|
|
Hardware:
CPU: Intel
Xeon 2.0 GHz
Memory: 1
GB PC2100 DDR registered ECC, dual channel
Hard drive: 37
GB SCSI for master node, 40 GB IDE for slave nodes
Motherboard: Tyan S2720 UGN/GN with Intel E7500 chipset
Network card: onboard
Intel Gigabit Ethernet controller 82554GC
Network Switch: Netgear GS516T 16-port Gigabit switch
KVM Switch: Aten CS1016 16-port
Software:
Operating System: RedHat Linux 7.3
Compiler: Intel
Fortran compiler 6.0 for Linux
MPI: LAM/MPI
6.5
Cost: 22K
Three-dimensional ZND
wave
Pressure contour
Mesh: 1.8 million
(120x120x128) mesh cells
Running time: 10
hours on Time Surfer II, May 2000
Refer to http://cfd.eng.ohio-state.edu/research/znd3d.htm.
Transverse Injection in a
Supersonic Flow
Vorticity
contour
Mesh: 5.5 million
(220x120x208) mesh cells
Running time: one
day on Time Surfer III, Dec 2002
Refer to http://cfd.eng.ohio-state.edu/research/injection/injection030103.htm.
Pulse Detonation Engine
Pressure Contour
Mesh: 6.4 million
(4000x1600) mesh cells
Running time: 88
hours for one cycle (roughly 83,000 steps of iteration) on Time Surfer III, Jan
2003
Refer to http://cfd.eng.ohio-state.edu/research/pde.htm.
Beowulf:
http://www.beowulf.org
Linux:
http://www.linux.org/
RedHat:
http://www.redhat.com
LAM/MPI:
http://www.lam-mpi.org/
MPICH:
http://www-unix.mcs.anl.gov/mpi/mpich/
METIS:
http://www-users.cs.umn.edu/~karypis/metis/
Intel Compiler:
http://www.intel.com/software/products/compilers/flin/
PGI CDK:
http://www.pgroup.com
Netperf:
http://www.netperf.org