A Scheduling Heuristic to Handle Local and Remote Memory in Cluster Computers

Abstract

In cluster computers, RAM memory is spread among the motherboards hosting the running applications. In these systems, it is common to constrain the memory address space of a given processor to the local motherboard. Constraining the system in this way is much cheaper than using a full-fledged shared memory implementation among motherboards. However, in this case, memory usage might widely differ among motherboards depending on the memory requirements of the applications running on each motherboard. In this context, if an application requires a huge quantity of RAM memory, the only feasible solution is to increase the amount of available memory in its local motherboard, even if the remaining ones are underused. Nevertheless, beyond a certain memory size, this memory budget increase becomes prohibitive. In this paper, we assume that the Remote Memory Access hardware used in a Hyper Transport based system allows applications to allocate the required memory from remote motherboards. We also analyze how the distribution of memory accesses among different memory locations (local or remote) impact on performance. Finally, an heuristic is devised to schedule local and remote memory among applications according to their requirements, and considering quality of service constraints.