Heuristics for work distribution of a homogeneous parallel dynamic programming scheme on heterogeneous systems

Dynamic Load Balancing on Dedicated Heterogeneous Systems

“…This approach usually implies at least a knowledge of said characteristics, such that the parallel program's tasks can be allocated according to the computational capacity of each processor. A more general approach can be obtained in the context of self-optimization strategies based on a run time model [4,9]. In this approach, an analytical model that parametrizes the architecture and the algorithm is instantiated for each specific case so as to optimize program execution.…”

“…This library facilitates the programmer the task of tailoring parallel code developed for homogeneous systems to heterogeneous ones [4]. The library has been implemented in a way that does not require changing any line of code in existing programs, thus minimizing code intrusion.…”

Processes Distribution of Homogeneous Parallel Linear Algebra Routines on Heterogeneous Clusters

“…In our previous study about work distribution for parallel dynamic programming [8], each node represented each of the T types of processors, but now it is necessary to manage each individual processor separately because for parallel linear algebra routines it is also necessary to obtain the best logical topology of the processes (a 2D mesh). Each node in the tree represents one possible solution, since this can be formed by any number of processes and more than one process can be assigned to a processor.…”

“…In a Backtracking or Branch-and-Bound scheme the LET (node) and M EET = min node∈evaluated nodes {GET (node)} are compared in each node to avoid working in the branches from this node when LET (node) ≥ M EET . In [8] five methods were introduced for parallel dynamic routines. Now, these methods have been adapted to the new situation presented for parallel linear algebra routines.…”

“…In this paper, we are faced with a heterogeneous scenario where, as well as the above decisions, new decisions must be taken: how many processes to generate and how many processes to assign to each processor (in the HoHo methodology, when one processor was chosen only one process was assigned to it [15]). Therefore, we propose an extension of the HoHo methodology, and we combine it with our previous approach of work distribution for a dynamic programming scheme [8].…”