LMDBdataset uses too much CPU #1331
-
|
Hello, I am using the OC20 dataset (in lmdb format) to train my own model, but I find it consume too much CPU. When I train my model on a single GPU, the CPU consumption is about 500%-550%, is it normal ? How can I improve it ? |
Beta Was this translation helpful? Give feedback.
Replies: 1 comment
-
|
Hi @MSJavaScript - can you provide some profiling here? How many dataset loading workers are you using? If the GPU utilization is high and all of the dataset loaders are running to keep it filled, this doesn't feel like an issue. Happy to follow up if you have specific notes on which sections of the code are unnecessarily expensive. Also, we switched to aselmdb format (now a default ASE DB format!) so please try that and see if the performance improves if low cpu utilization is important. |
Beta Was this translation helpful? Give feedback.
Hi @MSJavaScript - can you provide some profiling here? How many dataset loading workers are you using?
If the GPU utilization is high and all of the dataset loaders are running to keep it filled, this doesn't feel like an issue. Happy to follow up if you have specific notes on which sections of the code are unnecessarily expensive.
Also, we switched to aselmdb format (now a default ASE DB format!) so please try that and see if the performance improves if low cpu utilization is important.