As others have mentioned, Linpack benchmarks like those used in the Top 500 are largely irrelavent for measuring performance of climate models. Details of this
point are found in the recent
report from the National Academies of Science on Improving the Effectiveness of U.S. Climate Modeling. Massively parallel machines typically achieve 10% of their theoretical peak speed when running climate models while vecter parallel machines (NEC, Fugitsu) typically achieve 33% of their peak speed.
On a more theoretical note, Amdahl's law states that if a code has any serial portions, then the speedup limit is function of the code and not of the number of processors. For instance, if 0.1% (measured in single processor execution time) of the code does not parallelize, then regardless of how many cpu's you throw at the problem, the maximum speed up compared to a single cpu is 1000. Therefore fast VPP is the way to get more performace.
Then of course, there's the human resource issue. It's a real challenge to find skilled software engineers and then to convince them to accept low-paying government salaries.
I've spent six months with an O2000 suffering from irreproducable bus errors, segmentation faults, crashes and CPU hang-ups. SGI has not been able to fix it. It's a level of instablity orders beyond what I ever experienced with Cray (pre-SGI) or any other UNIX.
Slashdot Mirror
On a more theoretical note, Amdahl's law states that if a code has any serial portions, then the speedup limit is function of the code and not of the number of processors. For instance, if 0.1% (measured in single processor execution time) of the code does not parallelize, then regardless of how many cpu's you throw at the problem, the maximum speed up compared to a single cpu is 1000. Therefore fast VPP is the way to get more performace.
Then of course, there's the human resource issue. It's a real challenge to find skilled software engineers and then to convince them to accept low-paying government salaries.
This seems to be new
<a href="http://www.opera.com/linux/">
maybe it will run java.
I've spent six months with an O2000 suffering from irreproducable bus errors, segmentation faults, crashes and CPU hang-ups. SGI has not been able to fix it. It's a level of instablity orders beyond what I ever experienced with Cray (pre-SGI) or any other UNIX.