On 06/05/2012 02:31 PM, MORITA Kazutaka wrote: > If connect is the bottleneck, caching fds looks a simpler approach. I am not sure if fd pool will solve this problem, because we notice that obj_to_sheep() too occupy the CPU, which we use to get node location. These two time hoggers are linear with number of nodes in the cluster in recovery process. So I suspect that pull style is not that scalable. For a push style, we can get rid of get_obj_list(). But this change would be too radical, and should be considered at least for next release cycle. Thanks, Yuan |