potential memory leak when using snapshot
Hi, I've been trying to make a movie from volume renderings with Yt. The data is available as uniform numpy arrays, which are loaded using non-Yt infrastructure. For each movie frame, the following is done: Create Yt data from 3D numpy array rho dct = {"density":rho} ytdat = yt.load_uniform_grid(dct, sh, bbox=bbox) Create camera cam = ytdat.camera(c, L, W, N, self.transf, fields['density'], north_vector=[0,0,1], steady_north=True, sub_samples=5, log_fields=[True]) cam.transfer_function.map_to_colormap(self.lgmin,self.lgmax, scale=15.0, colormap='algae') Save movie frame as png: cam.snapshot(fn=name) Even with medium sized datasets (60^3), this runs out of memory after 30 frames or so. Some observations: Doing the same thing, but without the cam.snapshot(fn=name), the memory leak is gone. Creating the camera just once from the first dataset, and use cam.snapshot every frame does *not* increase memory usage. Of course, it then renders the same data over and over. Maybe I am doing something wrong. For example, is there a way to just replace the dataset for an existing camera? Any hints ? Best Regards, Wolfgang Kastaun.
Hi Wolfgang, This is indeed puzzling; I believe that what's happening is that the volume isn't created until you do the snapshot. I'm going to attempt to replicate your bug here and let you know. -Matt On Mon, Jun 22, 2015 at 6:24 AM, Wolfgang Kastaun <physik@fangwolg.de> wrote:
Hi,
I've been trying to make a movie from volume renderings with Yt. The data is available as uniform numpy arrays, which are loaded using non-Yt infrastructure. For each movie frame, the following is done:
Create Yt data from 3D numpy array rho dct = {"density":rho} ytdat = yt.load_uniform_grid(dct, sh, bbox=bbox)
Create camera cam = ytdat.camera(c, L, W, N, self.transf, fields['density'], north_vector=[0,0,1], steady_north=True, sub_samples=5, log_fields=[True])
cam.transfer_function.map_to_colormap(self.lgmin,self.lgmax, scale=15.0, colormap='algae')
Save movie frame as png: cam.snapshot(fn=name)
Even with medium sized datasets (60^3), this runs out of memory after 30 frames or so. Some observations:
Doing the same thing, but without the cam.snapshot(fn=name), the memory leak is gone.
Creating the camera just once from the first dataset, and use cam.snapshot every frame does *not* increase memory usage. Of course, it then renders the same data over and over.
Maybe I am doing something wrong. For example, is there a way to just replace the dataset for an existing camera?
Any hints ?
Best Regards, Wolfgang Kastaun.
_______________________________________________ yt-users mailing list yt-users@lists.spacepope.org http://lists.spacepope.org/listinfo.cgi/yt-users-spacepope.org
Hi Wolfgang, Sorry for the disappointing reply. I have attempted to identify the memory leak, and have come up short; I've got a few other things that I need to try first, but my *guess* is that there is a circular reference somewhere, or references being held to small numpy arrays. There's also the possibility that some of the memory is not being returned to the operating system, but I don't want to blame that just yet. A few quick fixes you might try would be to import gc at the top of your script, and at the end of each loop do: "del ytdat; gc.collect(2)". Alternately, if you dataset only changes in the data inside it, you can swap that in place by updating the "density" key of the dictionary you feed in to the load_uniform_grid call. -Matt On Mon, Jun 22, 2015 at 10:25 AM, Matthew Turk <matthewturk@gmail.com> wrote:
Hi Wolfgang,
This is indeed puzzling; I believe that what's happening is that the volume isn't created until you do the snapshot. I'm going to attempt to replicate your bug here and let you know.
-Matt
On Mon, Jun 22, 2015 at 6:24 AM, Wolfgang Kastaun <physik@fangwolg.de> wrote:
Hi,
I've been trying to make a movie from volume renderings with Yt. The data is available as uniform numpy arrays, which are loaded using non-Yt infrastructure. For each movie frame, the following is done:
Create Yt data from 3D numpy array rho dct = {"density":rho} ytdat = yt.load_uniform_grid(dct, sh, bbox=bbox)
Create camera cam = ytdat.camera(c, L, W, N, self.transf, fields['density'], north_vector=[0,0,1], steady_north=True, sub_samples=5, log_fields=[True])
cam.transfer_function.map_to_colormap(self.lgmin,self.lgmax, scale=15.0, colormap='algae')
Save movie frame as png: cam.snapshot(fn=name)
Even with medium sized datasets (60^3), this runs out of memory after 30 frames or so. Some observations:
Doing the same thing, but without the cam.snapshot(fn=name), the memory leak is gone.
Creating the camera just once from the first dataset, and use cam.snapshot every frame does *not* increase memory usage. Of course, it then renders the same data over and over.
Maybe I am doing something wrong. For example, is there a way to just replace the dataset for an existing camera?
Any hints ?
Best Regards, Wolfgang Kastaun.
_______________________________________________ yt-users mailing list yt-users@lists.spacepope.org http://lists.spacepope.org/listinfo.cgi/yt-users-spacepope.org
Hi Wolfgang, I take that back, I found the problem. Here's a pull request that fixes it: https://bitbucket.org/yt_analysis/yt/pull-request/1620 You can keep memory lower if you also gc.collect() it, but it won't grow out of control with this fix -- or it doesn't for me at least. -Matt On Mon, Jun 22, 2015 at 12:37 PM, Matthew Turk <matthewturk@gmail.com> wrote:
Hi Wolfgang,
Sorry for the disappointing reply. I have attempted to identify the memory leak, and have come up short; I've got a few other things that I need to try first, but my *guess* is that there is a circular reference somewhere, or references being held to small numpy arrays. There's also the possibility that some of the memory is not being returned to the operating system, but I don't want to blame that just yet.
A few quick fixes you might try would be to import gc at the top of your script, and at the end of each loop do: "del ytdat; gc.collect(2)". Alternately, if you dataset only changes in the data inside it, you can swap that in place by updating the "density" key of the dictionary you feed in to the load_uniform_grid call.
-Matt
On Mon, Jun 22, 2015 at 10:25 AM, Matthew Turk <matthewturk@gmail.com> wrote:
Hi Wolfgang,
This is indeed puzzling; I believe that what's happening is that the volume isn't created until you do the snapshot. I'm going to attempt to replicate your bug here and let you know.
-Matt
On Mon, Jun 22, 2015 at 6:24 AM, Wolfgang Kastaun <physik@fangwolg.de> wrote:
Hi,
I've been trying to make a movie from volume renderings with Yt. The data is available as uniform numpy arrays, which are loaded using non-Yt infrastructure. For each movie frame, the following is done:
Create Yt data from 3D numpy array rho dct = {"density":rho} ytdat = yt.load_uniform_grid(dct, sh, bbox=bbox)
Create camera cam = ytdat.camera(c, L, W, N, self.transf, fields['density'], north_vector=[0,0,1], steady_north=True, sub_samples=5, log_fields=[True])
cam.transfer_function.map_to_colormap(self.lgmin,self.lgmax, scale=15.0, colormap='algae')
Save movie frame as png: cam.snapshot(fn=name)
Even with medium sized datasets (60^3), this runs out of memory after 30 frames or so. Some observations:
Doing the same thing, but without the cam.snapshot(fn=name), the memory leak is gone.
Creating the camera just once from the first dataset, and use cam.snapshot every frame does *not* increase memory usage. Of course, it then renders the same data over and over.
Maybe I am doing something wrong. For example, is there a way to just replace the dataset for an existing camera?
Any hints ?
Best Regards, Wolfgang Kastaun.
_______________________________________________ yt-users mailing list yt-users@lists.spacepope.org http://lists.spacepope.org/listinfo.cgi/yt-users-spacepope.org
Thanks for the prompt reply! The fix works: the memory usage follows a sawtooth pattern, but there is no more swapping. When adding also the gc.collect(2), the memory usage stays constant. Note I did not apply the Bitbucket pull request directly because I did not know how and because it seemed to be against a different branch (I am using yt_analysis/yt). Instead, I manually changed the corresponding line. Btw, I also tried the other proposed solution
gc.collect(2)". Alternately, if you dataset only changes in the data inside it, you can swap that in place by updating the "density" key of the dictionary you feed in to the load_uniform_grid call.
That fixed the memory leak, too (without the patch), but it does not pick up the change. Instead, I get the first picture over and over again (and very fast, I think it does not render again). Do I have to tell the camera that the dataset changed? Wolfgang. On 06/22/2015 08:23 PM, Matthew Turk wrote:
Hi Wolfgang,
I take that back, I found the problem.
Here's a pull request that fixes it:
https://bitbucket.org/yt_analysis/yt/pull-request/1620
You can keep memory lower if you also gc.collect() it, but it won't grow out of control with this fix -- or it doesn't for me at least.
-Matt
On Mon, Jun 22, 2015 at 12:37 PM, Matthew Turk <matthewturk@gmail.com> wrote:
Hi Wolfgang,
Sorry for the disappointing reply. I have attempted to identify the memory leak, and have come up short; I've got a few other things that I need to try first, but my *guess* is that there is a circular reference somewhere, or references being held to small numpy arrays. There's also the possibility that some of the memory is not being returned to the operating system, but I don't want to blame that just yet.
A few quick fixes you might try would be to import gc at the top of your script, and at the end of each loop do: "del ytdat; gc.collect(2)". Alternately, if you dataset only changes in the data inside it, you can swap that in place by updating the "density" key of the dictionary you feed in to the load_uniform_grid call.
-Matt
On Mon, Jun 22, 2015 at 10:25 AM, Matthew Turk <matthewturk@gmail.com> wrote:
Hi Wolfgang,
This is indeed puzzling; I believe that what's happening is that the volume isn't created until you do the snapshot. I'm going to attempt to replicate your bug here and let you know.
-Matt
On Mon, Jun 22, 2015 at 6:24 AM, Wolfgang Kastaun <physik@fangwolg.de> wrote:
Hi,
I've been trying to make a movie from volume renderings with Yt. The data is available as uniform numpy arrays, which are loaded using non-Yt infrastructure. For each movie frame, the following is done:
Create Yt data from 3D numpy array rho dct = {"density":rho} ytdat = yt.load_uniform_grid(dct, sh, bbox=bbox)
Create camera cam = ytdat.camera(c, L, W, N, self.transf, fields['density'], north_vector=[0,0,1], steady_north=True, sub_samples=5, log_fields=[True])
cam.transfer_function.map_to_colormap(self.lgmin,self.lgmax, scale=15.0, colormap='algae')
Save movie frame as png: cam.snapshot(fn=name)
Even with medium sized datasets (60^3), this runs out of memory after 30 frames or so. Some observations:
Doing the same thing, but without the cam.snapshot(fn=name), the memory leak is gone.
Creating the camera just once from the first dataset, and use cam.snapshot every frame does *not* increase memory usage. Of course, it then renders the same data over and over.
Maybe I am doing something wrong. For example, is there a way to just replace the dataset for an existing camera?
Any hints ?
Best Regards, Wolfgang Kastaun.
_______________________________________________ yt-users mailing list yt-users@lists.spacepope.org http://lists.spacepope.org/listinfo.cgi/yt-users-spacepope.org
yt-users mailing list yt-users@lists.spacepope.org http://lists.spacepope.org/listinfo.cgi/yt-users-spacepope.org
participants (2)
-
Matthew Turk -
Wolfgang Kastaun