Remote Reference Protocol — PyTorch 2.7 documentation (original) (raw)

This note describes the design details of Remote Reference protocol and walks through message flows in different scenarios. Make sure you’re familiar with theDistributed RPC Framework before proceeding.

Background¶

RRef stands for Remote REFerence. It is a reference of an object which is located on the local or remote worker, and transparently handles reference counting under the hood. Conceptually, it can be considered as a distributed shared pointer. Applications can create an RRef by callingremote(). Each RRef is owned by the callee worker of the remote() call (i.e., owner) and can be used by multiple users. The owner stores the real data and keeps track of the global reference count. Every RRef can be uniquely identified by a global RRefId, which is assigned at the time of creation on the caller of theremote() call.

On the owner worker, there is only one OwnerRRef instance, which contains the real data, while on user workers, there can be as many UserRRefs as necessary, and UserRRef does not hold the data. All usage on the owner will retrieve the unique OwnerRRef instance using the globally unique RRefId. A UserRRef will be created when it is used as an argument or return value inrpc_sync(),rpc_async() orremote() invocation, and the owner will be notified according to update the reference count. An OwnerRRef and its data will be deleted when there is no UserRRef instances globally and there are no reference to the OwnerRRef on the owner as well.

Assumptions¶

RRef protocol is designed with the following assumptions.

• Transient Network Failures: The RRef design handles transient network failures by retrying messages. It cannot handle node crashes or permanent network partitions. When those incidents occur, the application should take down all workers, revert to the previous checkpoint, and resume training.
• Non-idempotent UDFs: We assume the user functions (UDF) provided torpc_sync(),rpc_async() orremote() are not idempotent and therefore cannot be retried. However, internal RRef control messages are idempotent and retried upon message failure.
• Out of Order Message Delivery: We do not assume message delivery order between any pair of nodes, because both sender and receiver are using multiple threads. There is no guarantee on which message will be processed first.

RRef Lifetime¶

The goal of the protocol is to delete an OwnerRRef at an appropriate time. The right time to delete an OwnerRRef is when there are no livingUserRRef instances and user code is not holding references to theOwnerRRef either. The tricky part is to determine if there are any livingUserRRef instances.

Design Reasoning¶

A user can get a UserRRef in three situations:

1. Receiving a UserRRef from the owner.
2. Receiving a UserRRef from another user.
3. Creating a new UserRRef owned by another worker.

Case 1 is the simplest where the owner passes its RRef to a user, where the owner calls rpc_sync(),rpc_async(), orremote() and uses its RRef as an argument. In this case a new UserRRef will be created on the user. As the owner is the caller, it can easily update its local reference count on the OwnerRRef.

The only requirement is that anyUserRRef must notify the owner upon destruction. Hence, we need the first guarantee:

G1. The owner will be notified when any UserRRef is deleted.

As messages might come delayed or out-of-order, we need one more guarantee to make sure the delete message is not processed too soon. If A sends a message to B that involves an RRef, we call the RRef on A (the parent RRef) and the RRef on B (the child RRef).

G2. Parent RRef will NOT be deleted until the child RRef is confirmed by the owner.

In cases 2 and 3, it is possible that the owner has only partial or no knowledge at all about the RRef fork graph. For example, an RRef could be constructed on a user, and before the owner receives any RPC call, the creator user might have already shared the RRef with other users, and those users could further share the RRef. One invariant is that the fork graph of any RRef is always a tree, because forking an RRef always creates a new UserRRef instance on the callee (except if the callee is the owner), and hence every RRef has a single parent.

The owner’s view on any UserRRef in the tree has three stages:

1. unknown -> 2) known -> 3) deleted.

The owner’s view of the entire tree keeps changing. The owner deletes itsOwnerRRef instance when it thinks there are no living UserRRefinstances, i.e., when OwnerRRef is deleted, all UserRRef instances could be either indeed deleted or unknown. The dangerous case is when some forks are unknown and others are deleted.

G2 trivially guarantees that no parent UserRRef can be deleted before the owner knows all of its children UserRRef instances. However, it is possible that the child UserRRef may be deleted before the owner knows its parent UserRRef.

Consider the following example, where the OwnerRRef forks to A, then A forks to Y, and Y forks to Z:

If all of Z’s messages, including the delete message, are processed by the owner before Y’s messages. the owner will learn of Z’s deletion before knowing Y exists. Nevertheless, this does not cause any problem. Because, at least one of Y’s ancestors will be alive (A) and it will prevent the owner from deleting the OwnerRRef. More specifically, if the owner does not know Y, A cannot be deleted due to G2, and the owner knows A since it is A’s parent.

Things get a little trickier if the RRef is created on a user:

OwnerRRef ^ | A -> Y -> Z

If Z calls to_here() on the UserRRef, the owner at least knows A when Z is deleted, because otherwise,to_here() wouldn’t finish. If Z does not callto_here(), it is possible that the owner receives all messages from Z before any message from A and Y. In this case, as the real data of the OwnerRRef has not been created yet, there is nothing to be deleted either. It is the same as Z does not exist at all. Hence, it’s still OK.

Implementation¶

G1 is implemented by sending out a delete message in UserRRefdestructor. To provide G2, the parent UserRRef is put into a context whenever it is forked, indexed by the new ForkId. The parent UserRRef is only removed from the context when it receives an acknowledgement message (ACK) from the child, and the child will only send out the ACK when it is confirmed by the owner.

Protocol Scenarios¶

Let’s now discuss how the above designs translate to the protocol in four scenarios.