I agree with your idea of having different metrics for physical and logical CPUs. My thought was that both could be reported, and then hyperthreading would be determined be checking if the two were equal or not. This would provide more information than my original boolean "hyperthreading" metric.
You are also right that this will affect the web frontend, but I think it is making it more accurate, not screwing it up. The reason I tried to create a patch was that the load per CPU was essentially half of the true value because ganglia thought there were 4 processors. So on the cluster view page, all the nodes looked blue and rarely changed color, even when jobs were running. But even without load/CPU, I still want my hardware info to be accurate. That's why I think two metrics for physcial and logical processors would be nice. --Rick On 25 Jul 2003, Jason A. Smith wrote: > I strongly agree with this. If we are running CPU intensive jobs on our > cluster, I want to know how loaded they are, which means their load > relative to the number of logical CPUs, not physical CPUs. > > Won't this also screw up the webfrontend, since it uses the load > relative to the number of CPUs reported to calculate the overall load on > the clusters and rank them in the grid view? It would make them look > like they are more loaded than they really are. > > If you really want this then I suggest adding a new metric for the > number of physical CPUs and leave this one as the number of logical > CPUs. > > ~Jason
