[vtk-developers] Performance tuning

[Ken Martin]

I've been doing some performance tuning in prep for VTK 5.0 and there is an
issue I'd like to discuss. One of the changes in VTK CVS that wasn't in VTK
44 is that the pipeline now holds reference counts going upstream and
downstream. If you have a reference to the source of a pipeline the entire
pipeline will be held. In VTK 4.4 the pipeline only held references going
upstream (although there were pointers going downstream as well). So in VTK
4.4 if you held a reference to a source then the source would stay around
but the rest of the downstream pipeline could be freed. So here is the deal,
having full reference counting upstream and downstream creates a lot of work
for the garbage collector in some cases, these performance bottlenecks can
be worked around using the Deferred collection and FastDelete methods but
that requires some knowledge and smarts on the developer's part. The
alternative is to go back to VTK 4.4 style single connected pipeline that
has a weak downstream reference. In most cases that speeds up the GC
significantly (because the reference graphs are much smaller). So here's the
summarized question, which would you prefer?

1) Developers must be smart about GC but the pipeline will be a doubly
linked list with strong references

2) The Developer can be more ignorant about GC but might be bitten by the
weak downstream reference in the pipeline

Currently 1) is what we have in CVS and 2) is what is in VTK 4.4 and
earlier. Note, in either case we will still be using the GC, the main
performance problem has been that the doubly linked list has created far
larger GC networks that VTK 4.4 ever had to deal with.

Thanks
Ken