4 Replies Latest reply on Oct 23, 2007 1:26 PM by Seth Langston

    Computer Performance Optimization

    Seth Langston
      I have a dual processor, dual core machine (Xeon 5160 @3.00 gb for each core on each processor) with 4 gb of ram, and dual 10,000 rpm hard drives. The loaner computer I was using was a Core2 duo @2.6 gb with 4 gb of ram and dual 10,000 rpm hard drives. I am finding that when I run the same problem it takes longer on this new machine. This COSOMS setup has pin connectors, node to surface no penetration contacts, and over 1.1 million degrees of freedom. I ran it on CW07 4.0 and it took about 2 hours and 51 minutes on the old machine. On the new machine, the same problem took 3 hours and 22 minutes.

      Just to check some things, I installed COSMOS 08 on each computer and ran the same problem again. COSMOS 08 was much faster (although it gave different results), solving the problem in 1 hour and 4 minutes on the old machine and 1 hour and 11 minutes on the new machine. I am still confused though because the newer/faster computer takes longer. Any ideas on what I can do to speed this machine up?
        • Computer Performance Optimization
          genexxer genexxer
          Simplify things. Try a cantilevered prismatic beam with a load on the free end and see what you get. Eliminate all the complexities and the high node count to get some quick comparisons. If the above bears out a fundamental about the processors' species may be revealed.(like the core2 duo is inherently more efficient for what you are doing).
            • Computer Performance Optimization
              Seth Langston
              I made a test part to run a comparison. It was just a part so there were no pin connectors or contact. I found similar results with the part taking 1 min 20 sec on the newer/faster machine but 1 min 18 sec on the older/slower machine. I am using the FFE solver on each. I might try the Direct Sparse and see what happens.
              • Computer Performance Optimization
                Valentin Leung
                The Xeon architecture needs higher latencies FB-Dimm to run, so itmay allow workstation to have more memory, but in other hand, it'sa little bit slower... But I don't think it's the only explanationto your problem.
                  • Computer Performance Optimization
                    Seth Langston
                    I ran some benchmarking software (SiSoft Sandra) and the FB-Dimm latencies were lower on the Xeon than on the Core2Duo. The Xeon had better scores in all areas on the Arithmetic, Multi-Core Efficiency, Memory Bandwidth, and Cache & Memory Bandwidth benchmarks than the C2D.

                    I ran another study using a shell mesh on an assembly made of surface models. It has 1511148 DOF with only bonded contact. Meshing this problem takes more time than running it. From what I can see, the mesher is not multi-threaded so the Xeon does perform a little better. It meshed the model in 1 hr 25 min where the C2D meshed the model in 1 hr 30 min. For run times the C2D was a little faster again, completing the analysis in 43 min 19 sec vs 43 min 49 sec on the Xeon. While the solver was running, the Xeon and the C2D had a CPU usage of around 65%. Occasionally the CPU usage (for both machines) would drop down to zero for about 30 seconds to a minute (this is when a convergence parameter was achieved).

                    Right now I am leaning towards returning the Xeon computer and replacing it with a much cheaper Core2Duo or maybe a Core2Quad.