mboehm7 opened a new pull request #927:
URL: https://github.com/apache/systemml/pull/927
This patch fixes remaining issues of incorrect results for colMins and
colMaxs over compressed matrix blocks with shared DDC1 dictionaries.
Specifically, if the individual column groups have only partial overlap,
the shared dictionary contains a superset of column group distinct
values. Since aggregation functions like min and max are executed only
over the dictionary (without touching the compressed data), it led to
incorrect results as we find extreme values that do not actually exist
in the column group.
Three alternatives approaches could solve this: (1) drop shared
dictionaries, (2) execute colMins and colMaxs over the compressed data,
or (3) refactor the double array dictionary into a proper class
hierarchy and maintain additional meta data for shared dictionaries. We
decided for (3) in order to keep predictable performance, irrespective
of shared dictionaries and because this class hierarchy allows for
further improvements of shared dictionaries between any subsets of
column groups.
Additionally, this fix also cleanups incorrect estimates of the
individual column groups (because getValueSize was used in the estimates
as a number of values, although it gave the size in bytes) as well as
some of the Class-layout size estimation tests.
----------------------------------------------------------------
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
For queries about this service, please contact Infrastructure at:
users@infra.apache.org
|