Lines Matching refs:TensorFlow
1 ## How to compile, use and configure RDMA-enabled TensorFlow
4 ```Do you wish to build TensorFlow with VERBS-RDMA support [y/N]```
23 The design is based on TensorFlow r1.0. An RDMA path is added between servers for tensor transfer (weights, gradients, etc). The existing GRPC path remains and is responsible for "administrative" tasks, such as setting up the RDMA path, exchanging computation graphs, etc.
25 During the server setup, an RDMA manager is created to manage low-level RDMA components such as RDMA channel and RDMA adapter, an RDMA rendezvous manager is created to oversee send/recv operations between servers. Following the distributed TensorFlow design philosophy, the send operation is passive, i.e. merely placing a tensor in the local out-going table. It is the receive operation that actually initiates the tensor transfer.
27 TensorFlow dynamically allocates memory for tensors that are to be sent or received. This causes difficulty for RDMA operations where pinned memory is required. Few remedies are possible:
30 3. Following HKUST research on the use of GPU direct, and their [GDR implementation](https://github.com/tensorflow/tensorflow/blob/master/tensorflow/contrib/gdr/README.md), there is a smart way to benefit from the TensorFlow allocation theme which is mostly pool based, i.e allocators pre-allocate a large memory block, and allocate the tensors from there. By attaching a custom Visitor to relevant allocators, we can do a single registration of the entire memory block, which zeros the registration overhead. Once the block is registered, each new tensor allocated will be at a registered address, which will allow us to do direct RDMA writes to it.
55 The tensor transfer process is initiated when the receiver requests a tensor. In code it is done by calling **Rendezvous::Recv()** or **Rendezvous::RecvAsync()**. The TensorFlow base implementation handles the case where the requested tensor is located on the same node. The more interesting case where the requested tensor is located on a remote node (receiver != sender) is to be handled in a derivation of the pure virtual **BaseRemoteRendezvous::RecvFromRemoteAsync()**. TensorFlow provides a default GRPC based implementation which comes in the vanilla version but suffers in scalability when running large models. Our RDMA based implementation presumes to be more scalable. HKUST's contrib GDR implementation is more scalable than GRPC, and less scalable than ours, only because we did our evolution based on it.
