|
|
@@ -20,6 +20,7 @@ TODO: improve!
|
|
|
* Power-based scheduling::
|
|
|
* Profiling::
|
|
|
* CUDA-specific optimizations::
|
|
|
+* Performance debugging::
|
|
|
@end menu
|
|
|
|
|
|
Simply encapsulating application kernels into tasks already permits to
|
|
|
@@ -289,3 +290,29 @@ StarPU already does appropriate calls for the CUBLAS library.
|
|
|
|
|
|
Unfortunately, some CUDA libraries do not have stream variants of
|
|
|
kernels. That will lower the potential for overlapping.
|
|
|
+
|
|
|
+@node Performance debugging
|
|
|
+@section Performance debugging
|
|
|
+
|
|
|
+To get an idea of what is happening, a lot of performance feedback is available,
|
|
|
+detailed in the next chapter. The various informations should be checked for.
|
|
|
+
|
|
|
+@itemize
|
|
|
+@item What does the Gantt diagram look like? (see @ref{Gantt diagram})
|
|
|
+@itemize
|
|
|
+ @item If it's mostly green (running tasks), then the machine is properly
|
|
|
+ utilized, and perhaps the codelets are just slow. Check their performance, see
|
|
|
+ @ref{Codelet performance}.
|
|
|
+ @item If it's mostly purple (FetchingInput), tasks keep waiting for data
|
|
|
+ transfers, do you perhaps have far more communication than computation? Did
|
|
|
+ you properly use CUDA streams to make sure communication can be
|
|
|
+ overlapped? Did you use data-locality aware schedulers to avoid transfers as
|
|
|
+ much as possible?
|
|
|
+ @item If it's mostly red (Blocked), tasks keep waiting for dependencies,
|
|
|
+ do you have enough parallelism? It might be a good idea to check what the DAG
|
|
|
+ looks like (see @ref{DAG}).
|
|
|
+ @item If only some workers are completely red (Blocked), for some reason the
|
|
|
+ scheduler didn't assign tasks to them. Perhaps the performance model is bogus,
|
|
|
+ check it (see @ref{Codelet performance}).
|
|
|
+@end itemize
|
|
|
+@end itemize
|
