Details
-
Improvement
-
Status: Closed
-
Major
-
Resolution: Fixed
-
None
-
None
Description
Improve performance of mini-batch preprocessor for images. May involve writing a new matrix aggregation function to support multi-dimensional arrays.
I have a 2 segment GP5 cluster set up:
- preprocessing 50k training rows from CIFAR-10 fits into 3 buffers and takes ~1 hour (buffer size of 24415 is reported in the summary file) – i.e. used NULL buffer size
- preprocessing 10k training rows from CIFAR-10 fits into 1 buffer and takes ~2 minutes
More info:
If I use `buffer_size=5000` it takes 979 sec
If I use `buffer_size=500` it takes 75 sec
So I think there is an issue with large buffer sizes