This article describes threading and using a gpgpu for parallelizing work.
From the article, he compares the slowdown when threads need to share a resource to something like two people having to share a piece of paper. Also, stuff like the overhead for creating a process and communication of a gpgpu back to the main cpu will also cause a slowdown. In addition, the work needs to be suitable to separation into pieces that may be worked on independently. He also mentions the driver and disk i/o may also be bottlenecks.
No comments:
Post a Comment