The results indicate a wide range of performance. The AMD 760 chipset shows the best performance, while the older PIII Serverworks LE chipset shows poor performance. Keep in mind the tests are not intended to show which platform is faster and that the Tyan/Athlon and Intel/Xeon do have faster memory. In addtion, the compiler optimizations are identical for all platforms. The tests were designed to show the effect of running two identical programs on the same node. This scenario is common in some cluster environments.
Although the PIII Serverworks performed poorly, this does not discount the use of this motherboard as a dual node. Keep in mind, that these tests are sequential applications running on a single node. In a parallel environment, communication overhead may alleviate the memory contention by spending time with communication issues. For instance, if a Fast Ethernet cluster were to use this motherboard, the SMP memory contention issue may not be as pronounced. If a higher performing low overhead networking is used, then the memory contention may become a bigger performance issue. The Athlon MP 760 chipset shows a clear advantage in memory contention issues. This type of cluster node should show good performance with all types of networking. Indeed, it is possible, given the range of performance, that a dual CPU node cluster of slower processors, but better SMP performance, may out perform a cluster built from faster processors with lesser SMP performance.
Similar to the first paper , using the Intel compiler lowers the efficiency. This effect is presumably due to the better efficiency of the code produced by the compiler and thus more dependance on the memory subsystem.
In these tests, we have not considered communication issues or the mix of different programs on the same node. These issues will be addressed in upcoming reports. Your results may vary depending on they type of application you run on your cluster.