|
|
@@ -10,6 +10,7 @@ w@c -*-texinfo-*-
|
|
|
* Installing a Binary Package::
|
|
|
* Installing from Source::
|
|
|
* Setting up Your Own Code::
|
|
|
+* Benchmarking StarPU::
|
|
|
@end menu
|
|
|
|
|
|
@node Installing a Binary Package
|
|
|
@@ -287,3 +288,48 @@ so:
|
|
|
@example
|
|
|
$ STARPU_NCUDA=2 ./application
|
|
|
@end example
|
|
|
+
|
|
|
+@node Benchmarking StarPU
|
|
|
+@section Benchmarking StarPU
|
|
|
+
|
|
|
+Some interesting benchmarks are installed among examples in
|
|
|
+@code{$prefix_dir/lib/starpu/examples/}. Make sure to try various
|
|
|
+schedulers, for instance STARPU_SCHED=dmda
|
|
|
+
|
|
|
+@menu
|
|
|
+* Task size overhead::
|
|
|
+* Data transfer latency::
|
|
|
+* Gemm::
|
|
|
+* Cholesky::
|
|
|
+* LU::
|
|
|
+@end menu
|
|
|
+
|
|
|
+@node Task size overhead
|
|
|
+@subsection Task size overhead
|
|
|
+
|
|
|
+This benchmark gives a glimpse into how big a size should be for StarPU overhead
|
|
|
+to be low enough. Run @code{tasks_size_overhead.sh}, it will generate a plot
|
|
|
+of the speedup of tasks of various sizes, depending on the number of CPUs being
|
|
|
+used.
|
|
|
+
|
|
|
+@node Data transfer latency
|
|
|
+@subsection Data transfer latency
|
|
|
+
|
|
|
+@code{local_pingpong} performs a ping-pong between the first two CUDA nodes, and
|
|
|
+prints the measured latency.
|
|
|
+
|
|
|
+@node Gemm
|
|
|
+@subsection Matrix-matrix multiplication
|
|
|
+
|
|
|
+@code{sgemm} and @code{dgemm} perform a blocked matrix-matrix
|
|
|
+multiplication using BLAS and cuBLAS. They output the obtained GFlops.
|
|
|
+
|
|
|
+@node Cholesky
|
|
|
+@subsection Cholesky factorization
|
|
|
+
|
|
|
+@code{cholesky*} perform a Cholesky factorization (single precision). They use different dependency primitives.
|
|
|
+
|
|
|
+@node LU
|
|
|
+@subsection LU factorization
|
|
|
+
|
|
|
+@code{lu*} perform an LU factorization. They use different dependency primitives.
|
Nathalie Furmento