creating pngs in background mode leaks memory

[doutriaux1]

A LOT!!!

for i in range(120):
   x.plot(data,"default","isofill")
   x.png("file_%i.png" % i)
   x.clear()

run and look at memory in top.

[doutriaux1]

@dlonie any idea why the memory could grow when bg=True ? The renWin is different obviously. Could it be VTK related?
Note: I tried to comment out hideGui/showGUI in def png but no luck

[alliepiper]

I'm a bit confused, it doesn't look like the plot call in the sample code is using background mode?

In any case, it doesn't look like VTK is holding onto anything in offscreen mode:

import gc
gc.enable()
gc.set_debug(gc.DEBUG_LEAK)

def runTest():
    import vtk
    sphere = vtk.vtkSphereSource()

    mapper = vtk.vtkPolyDataMapper()
    mapper.SetInputConnection(sphere.GetOutputPort())

    actor = vtk.vtkActor()
    actor.SetMapper(mapper)

    ren = vtk.vtkRenderer()
    ren.AddActor(actor)

    renWin = vtk.vtkRenderWindow()
    renWin.AddRenderer(ren)

    renWin.OffScreenRenderingOn()
    renWin.SetSize(300, 300)

    renWin.Render()

runTest()
del runTest

gc.collect()

leaks = {}

for i,g in enumerate(gc.garbage):
    t = type(g)
    print "Leak %d: %s @0x%x:\n   %s"%(i, t, id(g), str(g)[:70])
    leaks[t] = leaks.get(t, 0) + 1

for t in leaks:
    print "%7d: %s"%(leaks[t], t)

print len(gc.garbage),"leaks total."

$ python vtk_offscreen.py 
0 leaks total.

Looks like the leaks are on the VCS side of things.

[doutriaux1]

@dlonie I'm not quite sure what's going. All @durack1 and I noticed is that if you plot with vcs bg=1 then the process mem size keeps growing. And looking at the code it seems that the only differenc eis this:

    if self.bg:
        self.renWin.SetOffScreenRendering(True)
        self.renWin.SetSize(self.canvas.bgX,self.canvas.bgY)

But obviously there must be something else.

[durack1]

@doutriaux1 @dlonie for perspective, the loop generating these images has ~1700 steps and as it passed the 1000 mark yesterday it was using 50GB! of memory.. So it's not a small leak at all.. Each png being generated is around the 100KB mark, so around 170MB total as an output..

[doutriaux1]

@dlonie also it seems to be coming from the png function not the rendering function.
also a few more options are set on our renWin, not sure if it's matter e.g:

      self.renWin.SetWindowName("VCS Canvas %i" % self.canvas._canvas_id)
      self.renWin.SetAlphaBitPlanes(1)
      ## turning on Stencil for Labels on iso plots
      self.renWin.SetStencilCapable(1)
      ## turning off antialiasing by default
      ## mostly so that pngs are same accross platforms
      self.renWin.SetMultiSamples(0)

[alliepiper]

Those additional settings on the renderwindow shouldn't affect the offscreen rendering mechanism. Well, MultiSampling might, but only if it were non-zero.

What is the memory usage for the following scenarios?

• bg + plot + png
• bg + plot
• plot + png
• plot

Also, have you printed out the leaking objects and compared the results with a single run? The objects with increased leaks may point you to the issue.

[alliepiper]

The 50GB would make sense if the plot map were leaking -- that holds on to all of the plotted datasets as well as the intermediate filter outputs, which would add up very, very quickly.

[alliepiper]

Oh, and by plot map I mean the map returned by VTKPlots.plot -- I've noticed that it doesn't get freed in between calls to plot.

[doutriaux1]

Ok I'll run your runTest and print the leaked object in bg vs not in bg, that should point us in the right direction. Thanks.