Recently, I’ve been working on some projects for fun, trying out some things I hadn’t worked with before, such as profiling.
But after profiling my code, I found out that my average GPU activity is around 50%. Apparently, the code frequently hangs for a few hundred milliseconds on the dataloader process. I’ve tried a few things in the dataloader: increasing/decreasing the number of workers, setting pin-memory to true or false, but neither seems to really matter. I have an NVME drive, so the disk is not the problem either. I’ve concluded that the bottleneck must be the CPU.
Now, I’ve read that pre-processing the data might help, so that the dataloader doesn’t have to decode the images, for example, but I don’t really know how to go about this. I have around 2TB of NVME storage, and I’ve got a couple datasets on the disk (ImageNet and INaturalist are the two biggest ones), so I don’t suppose I’ll be able to store them on the disk uncompressed.
Is there anything I can do to lighten the load on the CPU during training so that I can take advantage of the 50% of the GPU that I’m not using at the moment?
You must log in or register to comment.
Might be cache misses or RAM access times. NVMe is mind boggling slow compared to L2 cache
Does your GPU util peak at 100%? If not, increase the batch size until it does (roughly). A couple ideas: do any transforms on the GPU or before starting your training job, have the CPU be solely responsible for loading images from disc. Increase the number of workers to the number of CPUs you have.
Which Api do you use ?
Number of workers should be the number of cpus. The preprocessing should be done in the dataset class not data loader.
I’ve dealt with similar issues in my own projects.
A couple of pointers :-
Use Image formats that are fast to decode for example BMP ( you can try converting all your images to BMP before you start training ) This will increase their size on disk but should reduce the CPU load. If you are doing any complex preprocessing on large images in your dataset class, try preprocessing images first and storing them to disk and loading those directly
These are just some general suggestions. It’d be more helpful if we knew more about your task so that we can offer more directed suggestions :)
Check your VRAM usage. If it’s above your dedicated VRAM, then you may be waiting for data to transfer from shared (system RAM) to your GPUs dedicated memory. You can also check the task manager while training to see the copy buffer usage and see if that’s constantly spiking throughout training.
As other have rightly pointed out, verify you’re using the Data Loader the right way. Ideally you need to create a custom dataset (in PyTorch terms) and apply all the transformations in this custom dataset. This might be helpful. Also, have you tried PyTorch Lightning?
CPU bottlenecks can be easily found by monitoring the CPU usage during training. If all of your cores are constantly at 100% your cpu might be too slow. If both the cpu and gpu are idle from time to time your storage could be the bottleneck.
To increase data loading performance, you could try out Nvidias Dali or FFCV which are both libraries optimized for that purpose. They replace some of the inefficient python code with highly optimised code. FFCV is quite nice but it requires you to convert your dataset into a specific format.