Investigate benchmark regressions

Registered by Anders Logg

The benchbot is showing quite a few regressions over the last years since DOLFIN 0.7.3. We need to investigate the reason for the regressions. The following regressions can be noted (and maybe some more):

1. 4 times slowdown in creating unit cube

2. slowdown in assembly

3. 3 times slowdown in mesh iteration

4. slowdown in jit compiler

5. gradual slowdown in refinement

6. dramatic slowdown in vector assignment

Blueprint information

Status:
Complete
Approver:
None
Priority:
High
Drafter:
None
Direction:
Needs approval
Assignee:
Johannes Ring
Definition:
Obsolete
Series goal:
Accepted for 1.1.x
Implementation:
Unknown
Milestone target:
None
Completed by
Garth Wells

Related branches

Sprints

Whiteboard

GNW: The vector assignment test is not a suitable test and the name is misleading. It tests the assignment of individual elements which will always be very slow and is only intended for debugging. The benchmark should test the function GenericVector::set for setting multiple elements at once.

AL: Sure. Just go ahead and add it. You can rename the current benchmark to something like indexassignment. I think we can have both (and you can ignore one of them if you don't think it is important ;-).

MER: I'm bumping this up -- would be good to have a look at it before releasing 1.0

AL: Too late for this now so I'm moving this to 1.1. Any fixes would require changes to existing algorithms and will likely introduce new bugs.

(?)

Work Items

This blueprint contains Public information 
Everyone can see this information.