Using Hardware Resource Allocation to Balance HPC Applications
Carlos Boneti, Roberto Gioiosa, Francisco J., Mateo Valero · InTech eBooks · 2010
In this chapter we present the problem of imbalance in HPC applications. In fact, some applications show an imbalanced behavior, i.e., some processes require more time to complete their computing phase while all the other processes are waiting at some synchronization point and cannot move forward. We show the reasons for imbalance and some examples where the application is imbalanced because of data distribution (NAS BTMZ), or because of the application's input (SIESTA). We also present the idea of using software controlled allocation of the hardware resources to perform load-balance of HPC applications. Experimental cases show how using a modified Linux kernel to control a processor capable to dynamically assign processor resources to running contexts (the IBM POWER5 in this case), reduces the application imbalance and, therefore, improves overall performance. The experiments performed show an improvement up to 18% for a widely used BT-MZ benchmark and up to 8.1% for a real application (SIESTA). These results do not require putting the burden of balancing the application on the programmer and are independent from the used programming model. In addition, we show cases where the application presents variable behavior. We discuss on why it motivates the use of automatic load-balancers based on software-controlled hardware resource allocation. From the case studies presented, it is possible to conclude that the hardware resource allocation in multithreaded processors is an important tool that allows to load-balance HPC applications, improving significantly their performance.