Suppressing independent loops in packing/unpacking loop nest to reduce message size for message-passing code
Presentation
Overview
Overview
Description
In this paper we experiment with two
optimization techniques we are considering
implementing in a parallelizing compiler that generates
parallel code for a distributed-memory system. We
have found that there are two problems that often arise
from the automatically generated message-passing
code: 1) messages contain redundant data, and 2) the
same data is sometimes transmitted to different
processors, yet the messages are repacked for each
processor.
Our experiments demonstrate that it is indeed
worthwhile suppressing the packing of redundant
information in a message. Not only did it improve
performance, but it allowed us to run the program on a
larger input size. We also discovered that it is not
worthwhile to suppress the repacking of the same
message. The reason is because the size of the
messages is a greater factor in the performance of a
message-passing program than the number of
instructions executed.
