Performance and Energy Analysis of the Iterative Solution of Sparse Linear Systems on Multicore and Manycore Architectures

Consumo energético de métodos iterativos para sistemas dispersos en procesadores gráficos

“…(For an analogous study with SP arithmetic, refer to [99].) For each architecture and matrix, we report the combination of core number and frequency that yields the shortest execution time per iteration, together with the corresponding net energy ("optimized w.r.t.…”

“…Las investigaciones del grupo High Performance Computing & Architectures de la Universitat Jaume I, en colaboración con miembros del Engineering & Mathematics Computing Lab del Karlsruhe Institute of Technology, identificaron a la CPU como un sumidero importante de energía durante la ejecución de métodos asíncronos de solución de sistemas lineales dispersos en servidores híbridos CPU-GPU [6,7].…”

“…We first review the study in [99], replacing the singleprecision (SP) arithmetic with double-precision (DP) experiments, as this is the norm in numerical linear algebra. In [99] we refrained from applying any manual hardware-aware optimizations to the code, but instead evaluated basic SP implementations of the CG method and SPMV, using either the baseline CSR (compressed sparse rows) or ELLPACK sparse matrix formats [129], and relying exclusively on the optimizations applied by the compiler. That scenario thus provided insights on the resource efficiency achieved when running complex numerical codes on large HPC facilities without applying hands-on optimization.…”

“…Except where stated otherwise, all tests were conducted with DP arithmetic. (We note that the experimentation in [99] was performed using SP arithmetic only.) Table 3.2 lists the main features of the hardware systems and compilers utilized in the experiments.…”

“…Despite the importance of energy consumption [163,165], few analyses of sparse linear algebra operations focus on this metric [157]. One particular source of energy inefficiency during the execution of an iterative solver [166] on a heterogeneous CPU-GPU server is that, when implemented via calls to the GPU kernels in CUBLAS/cuSPARSE, the CPU thread in control of the GPU repeatedly invokes fine-grain CUDA kernels of low cost and, therefore, short duration.…”