|
|
PDF SX-8 Data sheet ( Hoja de datos )
| Número de pieza | SX-8 | |
| Descripción | Cluster Documenation | |
| Fabricantes | NEC | |
| Logotipo |  |
|
Hay una vista previa y un enlace de descarga de SX-8 (archivo pdf) en la parte inferior de esta página. Total 25 Páginas | ||
|
No Preview Available !
www.DataSheet4
NEC SX-8 Cluster Documentation
Holger Berger
January 18, 2008
1 page  
www.DataSheet4
Chapter 1
Introduction
After HLRS offered access to SX-6 nodes for a year, the SX-8 system is now in operation. SX-8 and
SX-6 are sharing a lot of characteristics, so SX-6 users should feel comfortable with SX-8 from the
beginning. SX-8 offers with 72 nodes with 8 fast vector CPUs a so far unknown level of performance
and throughput to HLRS users.
1.1 Hardware and Architecture
The system consists of two scalar frontend system, and 72 SX-8 nodes. The scalar frontend systems
is a 32CPU Itanium2 system called asama, with a very large memory of 240 GB and a scalar frontend
systems called a1 with 32CPU Itanium2 and 512GB of memory. Each CPU has a large 6MB L3
cache.
The frontend systems and the backend systems share fast filesystems. A total of two 9TB filesys-
tems is used for user homes. One 80TB filesystem contain workspace which can be used during
runtime of jobs. The workspace filesystem can sustain about 10GB/s in aggregate, for a single node
about 600MB/s are possible.
The backend systems are 72 SX-8 vector nodes, each having 8 CPUs of 16 Gflops peak (2Ghz).
Each node has 128 GB, about 124 GB are usable for applications.
The vector systems have a fast interconnect called IXS, which is a central crossbar switch. Each
node can send and receive with 16 GB/s in each direction. You can expect an MPI latency around 5
microseconds for small messages. Each SX-8 CPU can access the main memory with 64 GB/s.
Features:
• Cluster of 72 SX-8 nodes
• Frontend is two times 32way NEC TX-7
• Batchsystem: NQSII
• Operating System: TX7: SUSE SLES9, SX8: SUPER-UX 15.1
1.2 Access
All former SX-6 users are registered for SX-8 and new frontend TX-7 a1.hww.de. Password was
copied from crossi.hww.de. The a1.hww.de is integrated in the HLRS LDAP concept, so if you
change the password on any other LDAP machine, it is reflected on a1.hww.de as well.
5
5 Page 
www.DataSheet4
2.3. CREATING BATCH JOBS
11
rm *.data
# ... or whatever your files are called
This example shows a 4 CPU jobs, where two MPI processes spawn two threads each using
OpenMP. The total CPU time is below 2 hours, and the jobs does not use more than 10 GB of memory.
It is running on one node. The Jobname is for your convenience only, the account code should always
be specified (otherwise, your jobs can not start!). After 2 hours, the job will be stopped due to elapsed
time limit.
If you do not specify a CPU number, 1 CPU is assumed. Please note: The number of CPUs spec-
ified here has a meaning. It is used to find free resources by the batch system, and it is used to assign
CPUs to your processes during runtime. But you still have to specify F RSVTASK, OMP NUM THREAD
or mpirun -np to really get this number of CPUs in the mix of processes and threads you want them
to have.
A serial job can use up to 64 GB, and parallel job up to 124 GB.
2.3.2 Multi-node jobs
#!/usr/bin/ksh
#PBS -q multi
# queue: dq for <=8 CPUs
#PBS -T mpisx
# Job type: mpisx for MPI
#PBS -l cpunum_job=8
# cpus per Node
!!!!!!!!!!!!!!!
#PBS -b 4
# number of nodes
#PBS -l elapstim_req=02:00:00 # max wallclock time
<!!!!!!!! NEW
#PBS -l cputim_job=08:00:00 # max accumulated cputime per NODE
#PBS -l cputim_prc=07:55:00 # max accumulated cputime per process
#PBS -l memsz_job=10gb
# memory per node
#PBS -A <account code>
# Your Account code, see login message (without <>!)
#PBS -j o
# join stdout/stderr
#PBS -N MyJob
# job name
#PBS -M jo@user
# you should always specify your email
SCR=‘ws_allocate Run1 2‘ # workspace for 2 days
cd $SCR
export OMP_NUM_THREADS=8
export MPIPROGINF=YES
export F_FILEINF=YES
export MPIMULTITASKMIX=YES
MPIEXPORT="OMP_NUM_THREADS F_FILEINF"
export MPIEXPORT
mpirun -nn 4 -nnp 1 ./mycode
cp outfile $HOME/code
# please be so kind and clean up your workspace from files no longer needed
This is a multinode example. Please note the directive -T mpisx, which is mandantory for
multinode jobs!
This job used 4 nodes. It uses the first 4 nodes which are assigned by the batch system (-nn 4),
one mpi process per node (-nnp 1). Each of them creates 8 OpenMP threads. Please note the usage of
11 Page | ||
| Páginas | Total 25 Páginas | |
| PDF Descargar | [ Datasheet SX-8.PDF ] | |
Hoja de datos destacado
| Número de pieza | Descripción | Fabricantes |
| SX-205 | Stereo Receiver Service Manual |  Pioneer |
| SX-255R | Stereo Receiver Service Manual | Pioneer |
| SX-4001A | SX-4001A |  EIC discrete Semiconductors |
| SX-8 | Cluster Documenation | NEC |
| Número de pieza | Descripción | Fabricantes |
| SLA6805M | High Voltage 3 phase Motor Driver IC. |
 Sanken |
| SDC1742 | 12- and 14-Bit Hybrid Synchro / Resolver-to-Digital Converters. |
 Analog Devices |
|
DataSheet.es es una pagina web que funciona como un repositorio de manuales o hoja de datos de muchos de los productos más populares, |
| DataSheet.es | 2020 | Privacy Policy | Contacto | Buscar |