Multi-SPMD Programming Model with YML and XcalableMP
229
Fig. 9 Execution timeline of the BGJ workflow application with 8 × 8 blocks
2048 processes for 8 × 8 blocks, more than 256 processes for 16 × 16 processes,
show poor performance.
Figure 9 shows the execution timeline (from left to right) of the BGJ workflow
application with 8 × 8 blocks. As shown in the figure, at the first step, the task
of inversion (B = A −1 ) must be executed solely since the other tasks on the
first step use the result of the inversion. After the second step, some of the matrix
calculations such as A = A × B, C = −(B × A), C = C − (B × A) on the kth
step and the inversion on k + 1th step can be overlapped. For other programming
models such as flat-MPI, it is not easy to execute tasks or functions on different
steps simultaneously. On the other hand, the mSPMD programming model and
our programming environment allow application developers to describe this sort
of applications easily.
6 Eigen Solver on the mSPMD Programming Model
In this section, as a use case of the mSPMD programming model, we introduce an
eigen solver implemented on the mSPMD programming model.
Précédent

- 233/265

Suivant