Re: Question about displaying process memory allocation on Linux

[Mark Rivers <rivers@cars.uchicago.edu>]

Hi Mark,


> but do you have a suggested place where you might suggest putting NDArrayPool::emptyFreeList() so that memory will always be flushed, when it can be?


That's a tricky question.  The point of the free list is that it should be more efficient than calling malloc() and free() for the large NDArrays each time a new one is allocated and freed.  If the free list were regularly emptied it would defeat this optimization.  That being said I have not actually benchmarked the time to use NDArrayPool->alloc vs calling new NDArray(ndims, dims, dataType, 0, NULL).


> And perhaps a place where it can be done automagically when memory is about to be exceeded so that AD just starts to dump frames instead of crashing because there is no free memory left?


If by "no free memory left" you mean exceeding maxMemorySize then that is already the case, and has been for a long time.  If NDArrayPool->alloc() would exceed the maxMemorySize then it deletes NDArrays from the free list until enough memory is available (if possible).


However, prior to R3-3 each driver and plugin had its own NDArrayPool, and so it was difficult to use maxMemorySize effectively.  It is the total memory in the IOC that matters, but previously one had to set the maxMemorySize for each driver and plugin.  In R3-3 all plugins get their NDArrays from the driver's NDArrayPool, so a single setting of maxMemorySize is all that is needed, and it controls that driver and all of its plugins.


Mark



________________________________
From: Mark Engbretson <engbretson@aps.anl.gov>
Sent: Tuesday, June 26, 2018 9:03 AM
To: Mark Rivers; tech-talk@aps.anl.gov
Subject: RE: Question about displaying process memory allocation on Linux


Not an answer to your Linux issues . . . sorry . . . . but do you have a suggested place where you might suggest putting NDArrayPool::emptyFreeList() so that memory will always be flushed, when it can be? And perhaps a place where it can be done automagically when memory is about to be exceeded so that AD just starts to dump frames instead of crashing because there is no free memory left?





From: tech-talk-bounces@aps.anl.gov <tech-talk-bounces@aps.anl.gov> On Behalf Of Mark Rivers
Sent: Tuesday, June 26, 2018 8:52 AM
To: tech-talk@aps.anl.gov
Subject: Question about displaying process memory allocation on Linux



Folks,



In areaDetector the NDArrays are allocated from a free list in NDArrayPool.  Previously once a large number of arrays had been allocated on the free list there was no way to free that memory back to the OS without restarting the IOC.  I have now added an NDArrayPool::emptyFreeList() method and an EmptyFreeList record that results in a call to that method.



When I test on Windows and I look at the Commit Size for the simDetector process in Task Manager I see exactly what I expect.  If I have large queues and the detector is acquiring faster than the plugins can process the free list contains NDArrays totaling over 3GB.   The Commit Size is ~3.2 GB.  When I process the EmptyFreeList record the CommitSize immediately drops to 0.2 GB.



I am trying to see if I see the same thing on Linux, using the ps –o trs,vsz,drs command.



This is what I see when I first start the simDetector IOC, before I collect any images.  Note that the virtual memory size (VSZ) is already 5GB.  This seems strange.  Can anyone explain?

corvette:dxpSITORO/iocBoot/iocFalconX1>ps -p 115038 -o trs,vsz,drs

TRS    VSZ   DRS

15838 5080984 5065145



This is what I see after collecting enough images that there is a total of 3GB of NDArrays allocated.  VSZ has increased by 3GB (from 5GB to 8GB), which is what I expect.

corvette:dxpSITORO/iocBoot/iocFalconX1>ps -p 115038 -o trs,vsz,drs

TRS    VSZ   DRS

15838 8171424 8155585



This is what I see when acquisition has stopped and all plugins have competed.  The NDArrays are now all on the free list, VSZ remains at 8 GB, which is what I expect.

corvette:dxpSITORO/iocBoot/iocFalconX1>ps -p 115038 -o trs,vsz,drs

TRS    VSZ   DRS

15838 8171424 8155585



This is what I see after I process the EmptyFreeList record.  Note that VSZ has only decreased by about 0.2GB, not by 3GB which I would have expected.  Can anyone explain?

corvette:dxpSITORO/iocBoot/iocFalconX1>ps -p 115038 -o trs,vsz,drs

TRS    VSZ   DRS

15838 7974816 7958977



Is there another statistic I can look at to see that my process has indeed returned all 3GB to the OS?



Thanks,

Mark