|
|
Computational Resources
|
Computational Resources at the Center
| Computer | Memory | Hostname |
1 SGI Power Challenge Server
(4 CPUs, 90 MHz, R8000)
| 2 GB RAM, 2 GB Disk | powerc |
| 4 Alphastation 250 4/266 series Workstations | 192 MB RAM, 1.44 MB Floppy,
600 MB CD-ROM, 1.05 GB SCSI Disk
| laotze, socrates, buddha, confucius |
| 1 SGI Indigo 2 Extreme Workstation, 250 MHz | 64 MB RAM, 2 GB Disk | extreme |
| 2 SGI Indy Workstations, 100 MHz, R4600PC | 32 MB RAM, 535
MB Disk | indya, indyb |
To use the CCEM printer upstairs, login to one of the ccem computers and do
lpr -P hp4plus filename Matlab user please first save your figure as a
postscript file, then print.
Local Computers in Prof. Chew's Group
These are the computers everybody in the group should be able to access.
If you have problem,
please let me know.
| Computer | Memory | Hostname |
Ultra Sparc 143 MHz
Solaris 2.5.1 | 192 MB RAM | gspark |
Sparc 10
SunOS 4.1.3 | 128 MB RAM | espark |
2 Sparc 2
SunOS 4.1.3 | 64 MB RAM | sunchew, csaprk |
Sparc Classic
SunOS 4.1.3_U1 | 48 MB RAM | fspark |
Sparc Classic
SunOS 4.1.3_U1 | 24 MB RAM | dspark |
2 Sparc IPC
SunOS 4.1.3 | 24 MB RAM | aspark, bspark |
2 Pentium 100 MHz
Linux--Slackware/Windows 95 | 64 MB RAM | empc10, empc16 |
3 Pentium 166 MHz
Linux--Slackware/Windows 95 | 32 MB RAM | empc01, empc02, empc07 |
Please note that your passwd on the local computers is different from the one
on CCEM machines. The HP printer in 377 can be accessed by command
lw -P lp600 filename.ps
The other printer in 379 can be accessed by command
lw -P lp300 filename.ps
The color printer in 379 can only be accessed from dspark. You can do
the following to print filename.ps.
lw -P clw filename.ps
Others:
IMSL library and the Numerical Recipes codes are located under /mnts drectory.
The license for IslandDraw and Matlab is limitted. Please quit when
you are finished using them.
Power Challenge User Guide
1. Overall
The following is some hints based on my past usage experiences and might
contain error. For better results, please use on-line man page and
the SGI insight tool. Some manuals might also be available outside
our group, because the computer is very popular.
2. Debug Single CPU Job First
Make sure that the code is working before parallelizing it,
because it is much more difficult to debug in parallel mode.
Parallelizing your code could also introduce bugs or error due to
data dependence.
3. Where to Parallelize
Always profile your codes first to find where the cpu time goes.
Never parallelize the whole code, but parallelize only the part
which is cpu intensive.
4. How to Parallelize
Parallelize your code gradually and look at the difference in cpu time
before and after parallelization by using
following codes.
real tarray(2),time1,time2,etime
time1=etime(tarray)
time2=etime(tarray)
print*,'Elapsed Time =',time2-time1
If the gain is not significant, do not parallelize that part. By significant
I mean the cpu time of the parallel job (mp job) should be only slightly
larger than one quarter of the cpu time of the single thread job (sp job).
5. MP Command - DOACROSS
DOACROSS parallelizes do-loops. In my feeling, the parallelization
should be of coarse grain, because the computer does not have many cpus.
A high volume of scheduling would increase the overhead.
Use and declare as many LOCAL variables as possible instead of
using SHAREd variables.
6. Data Independence
Data dependence might give you wrong results.
Make sure you do not have data dependence by comparing the results
while gradually parallelizng your code.
7. Fortran Compiling
Always use selection -O3 for software pipelining (SWP). There could be
accuracy compromise. The selection -mips4 is set by default.
% f77 -O3 mycode.f
For your mp commands to take effect, use -mp selection.
% f77 -O3 -mp mycode.f
8. PFA - Power Fortran Accelerator
Use -mp selection only.
Never use -pfa selection. PFA usually
wastes the cpu time. Parallel job using PFA is often much
slower than your single process job.
9. Running Your Job
48x4 hours are set for extremely large jobs. For testing,
parallel jobs should not last for more than 1x4 hours and smaller
test case should be used. Please note that a 3D FD code of size
80x80x80x3 (number of unknowns) only takes less than half
an hour using parallel mode.
10. Environment Setting
Default is four cpus. To use fewer cpus do the following.
% setenv MP_SET_NUMTHREADS 1 .......Use one cpu only.
% setenv MP_SET_NUMTHREADS 2 .......Use two cpus.
11. Moral Issue
Please notice that the machine is shared by a large group and
try to respect other people's job. Single cpu job should be run
on other computers unless the job requires large memory. Please
also notice that too many jobs running simultaneously might cause
the computer to crash, thus causing inconvenience.
Last updated on January 9, 1997.
|
