Join GitHub today
GitHub is home to over 40 million developers working together to host and review code, manage projects, and build software together.Sign up
added thread lock acquire and release to OnDemandLinkSequence, to allow multithreading #477
This patch, inspired by this post, appears to have fixed the multithreading issue. The way to use is
@PantelisElinas the speed up is significant on CPU; let's try this on
Yes, the training time per epoch on
Before the fix, an error was thrown
the fix looks good and I have not come across the exception we used to get. Also, here are the timing results from our workstation.
Times are per epoch on CPU and GPU using multi-threading. The machine is Intel i9-7960X 2 2.80GHz with NVIDIA GeForce RTX 2080.
GPU utilisation for the above experiments was 10-15% (very low) but we already know that with GraphSAGE we have a bandwidth problem sending the node features to the GPU. In other words, there is little benefit using the GPU, in fact, the extra overhead of having to send data to the GPU hurts performance.