I'm trying to start multiple processes (10 for example) of learning with tensorflow 2. I'm still using Session and so on multiple tf.compat.v1 in all my codebase.
When I'm running with CPU, processes take each around 500mo of CPU memory. htop output :
When I'm running with GPU, processes take each much more CPU memory (like 3Go each) and almost the same (more in reality) GPU memory. nvtop output (GPU mem left, CPU (HOST) mem right) :
I can reduce GPU memory process fingerprint by using environment variable TF_CUDNN_USE_AUTOTUNE=0 (1.5Go GPU, not more than 3Go CPU). But it's still much more memory consumption than running process on CPU only. I tried a lot of thing like TF_GPU_ALLOCATOR=cuda_malloc_async with a tf nightly release, but it's still the same. This cause OOM errors if I would like to keep 10 processes on GPU like on CPU.
I found memory fragmentation may be a hint, by profiling a single process. You can find screenshots here.
TL;DR
When running tf process on CPU only, it uses some memory (comparable to data size). When running the same tf process on GPU only, it uses much more memory (~x16 without any tensorflow optimization).
I would like to know what can cause a huge difference of memory usage like this, and how to prevent it. Even how to fix it.
FYI -> Current setup : tf 2.6, cuda 11.4 (or 11.2 or 11.1 or 11.0), ubuntu 20.04, nvidia driver 370
EDIT : I tried to convert my tensorflow / tflearn code to pytorch. I have the same behaviour (low memory on CPU, and everything explode when running on GPU)
EDIT2 : Some of memory allocated on GPU should be for CUDA runtime. On pytorch. I have 300mo memory allocated on CPU run. I have 2go of GPU memory and almost 5go of CPU memory used when running on GPU. May the main problem is the CPU/system memory allocated for this process when I'm running on GPU, since it seems that CUDA runtime can take almost 2go of GPU mem (this is huge...). It looks like related to CUDA initialization.
EDIT3 : This is definitely an issue with CUDA. Even if I try to create a 1,1 tensor with pytorch, it takes 2go of GPU and almost 5go of CPU memory. It can be explain because pytorch is loading a huge number of kernels to memory; even if the main program isn't using them.
Related
This is a total newbie question but I've been searching for a couple days and cannot find the answer.
I am using cupy to allocate a large array of doubles (circa 655k rows x 4k columns ) which is about 16Gb in ram. I'm running on p2.8xlarge (the aws instance that claims to have 96GB of GPU ram and 8 GPUs), but when I allocate the array it gives me out of memory error.
Is this happening becaues the 96GB of ram is split into 8x12 GB lots that are only accessible to each GPU? Is there no concept of pooling the GPU ram across the GPUs (like regular ram in multiple CPU situation) ?
From playing around with it a fair bit, I think the answer is no, you cannot pool memory across GPUs. You can move data back and forth between GPUs and CPU but there's no concept of unified GPU ram accessible to all GPUs
When running a TensorFlow job I sometimes get a non-fatal error that says GPU memory exceeded, and then I see the "Shared memory GPU usage" go up on the Performance Monitor on Windows 10.
How does TensorFlow achieve this? I have looked at CUDA documentation and not found a reference to the Dedicated and Shared concepts used in the Performance Monitor. There is a Shared Memory concept in CUDA but I think it is something on the device, not the RAM I see in the Performance Monitor, which is allocated by the BIOS from CPU RAM.
Note: A similar question was asked but not answered by another poster.
Shared memory in windows 10 does not refer to the same concept as cuda shared memory (or local memory in opencl), it refers to host accessible/allocated memory from the GPU. For integrated graphics processing host and device memory is usually the same as shared thanks to both the cpu and gpu being located on the same die and being able to access the same ram. For dedicated graphics with their own memory, this is separate memory allocated on the host side for use by the GPU.
Shared memory for compute APIs such as through GLSL compute shaders, or Nvidia CUDA kernels refer to a programmer managed cache layer (some times refereed to as "scratch pad memory") which on Nvidia devices, exists per SM, and can only be accessed by a single SM and is usually between 32kB to 96kB per SM. Its purpose is to speed up memory access to data which is used often.
If you see and increase shared memory used in Tensorflow, you have a dedicated graphics card, and you are experiencing "GPU memory exceeded" it most likely means you are using too much memory on the GPU itself, so it is trying to allocate memory from elsewhere (IE from system RAM). This potentially can make your program much slower as the bandwidth and latency will be much worse on non device memory for a dedicated graphics card.
I think I figured this out by accident. The "Shared GPU Memory" reported by Windows 10 Task Manager Performance tab does get used, if there are multiple processes hitting the GPU simultaneously. I discovered this by writing a Python programming that used multiprocessing to queue up multiple GPU tasks, and I saw the "Shared GPU memory" start filling up. This is the only way I've seen it happen.
So it is only for queueing tasks. Each individual task is still limited to the onboard DRAM minus whatever is permanently allocated to actual graphics processing, which seems to be around 1GB.
Linux: Ubuntu 16.04.3 LTS (GNU/Linux 4.10.0-38-generic x86_64)
Tensorflow: compile from source, 1.4
GPU: 4xP100
I am trying the new released object detection tutorial training program.
I noticed that there is big difference when I set CUDA_VISIBLE_DEVICES to different value. Specifically, when it is set to "gpu:0", the gpu util is
quite high like 80%-90%, but when I set it to other gpu devices, such as
gpu:1, gpu:2 etc. The gpu util is very low between 10%-30%.
As for the training speed, it seems to be roughly the same, much faster than that when using CPU only.
I just curious how this happens.
As this answer mentions GPU-Util is a measure of usage/business of the computation of each GPU.
I'm not an expert, but from my experience GPU 0 is generally where most of your processes run by default. CUDA_VISIBLE_DEVICES sets the GPUs seen by the processes you run on that bash. Therefore, by setting CUDA_VISIBLE_DEVICES to gpu:1/2 you are making it to run on less busy GPUs.
Moreover, you only reported 1 value, in theory you should have one per GPU; there is the possibility you were only looking at GPU-util for GPU-0 which would of course decrease if you are not using.
When a Theano script tries to obtain more GPU memory than currently available, it immediately crashes:
Error allocating 26,214,400 bytes of device memory (out of memory).
Driver report 19,365,888 bytes free and 1,073,414,144 bytes total
Is there any way to configure Theano so that it doesn't directly crash when a GPU memory allocation fails? E.g., it could instead retry every X seconds, and give up after Y tries.
(One use case: I have several Theano scripts running on the same GPU and using 90% of its memory only 5% of the time)
Most of the benchmarks for gpu performance and load testing are graphics related. Is there any benchmark that is computationally intensive but not graphics related ? I am using
DELL XPS 15 laptop,
nvidia GT 525M graphics card,
Ubuntu 11.04 with bumblebee installed.
I want to load test my system to come up with a max load the graphics cards can handle. Are there any non-graphics benchmarks for gpu ?
What exactly do you want to measure?
To measure GFLOPS on the card just write a simple Kernel in Cuda (or OpenCL).
If you have never written anything in CUDA let me know and i can post something for you.
If your application is not computing intensive (take a look at a roofline paper) then I/O will be the bottleneck. Getting data from global (card) memory to the processor takes 100's of cycles.
On the other hand if your application IS compute intensive then just time it and calculate how many bytes you process per second. In order to hit the maximum GFLOPS (your card can do 230) you need many FLOPs per memory access, so that the processors are busy and not stalling for memory and switching threads.