Details
-
Sub-task
-
Status: Closed
-
Major
-
Resolution: Resolved
-
None
-
None
-
None
Description
In preparation of these experiments, it might be very useful to extend the parameter server and SystemML -stats}}tool to keep track where we spent how much time (e.g., time in compute gradients, aggregation, copying of matrices, slicing of batches). This would allow us to understand the bottlenecks better and address them in a systematic manner. Please have a look into {{org.apache.sysml.utils.Statistics. We could show a few summary statistics for the parameter server if at least one paramserv function was executed.