ode4j
ode4j copied to clipboard
tzaeschke
Garbage collection - step()
As mentioned in issue #35, ode4j appears to allocate significant amounts of memory in some situations/systems ("3-4 MB/s").
The problem may be in the stepper which allocates large arrays for the solver during each step.
A solution could involve array pooling.
I implemented an experimental hack which can be enabled with:
DxQuickStep.REUSE_OBJECTS = true;
This reuses a number of objects, including the large collision used in the stepper.
Could someone have a look whether it improves the situation on Android? We could apply the same approach to other places in the code.
It looks like some improvement has been recently made in ODE too - see https://bitbucket.org/odedevs/ode/commits/506a3b3b93c2ef0da3e5030bf12e408aef59c2cd
Yes, they seem to avoid allocating Ad, even if that is much smaller than other matrices (iMJ is 12 times bigger).
What I did in this patch is really just a hack and should definitely be extended to the matrix solver as well.
I also looked a few days ago at porting ODE changes to ode4j. But quickstep has changed considerably, so I suppose I need a week off for that. Maybe over the upcoming holidays.
Has anyone seen a measurably reduced GC load from my initial patch?