rref.rst.txt

Remote Reference Protocol

This note describes the design details of Remote Reference protocol and walks through message flows in different scenarios. Make sure you're familiar with the :ref:`distributed-rpc-framework` before proceeding.

Background

RRef stands for Remote REFerence. It is a reference of an object which is located on the local or a 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 calling :meth:`~torch.distributed.rpc.remote`. Each RRef is owned by the callee worker of the :meth:`~torch.distributed.rpc.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 the :meth:`~torch.distributed.rpc.remote` 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 in :meth:`~torch.distributed.rpc.rpc_sync`, :meth:`~torch.distributed.rpc.rpc_async` or :meth:`~torch.distributed.rpc.remote` 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 aims to handle transient network failures by retrying messages. Node crashes or permanent network partition is beyond the scope. When those incidents occur, the application may take down all workers, revert to the previous checkpoint, and resume training.
• Non-idempotent UDFs: We assume the user functions (UDF) provided to :meth:`~torch.distributed.rpc.rpc_sync`, :meth:`~torch.distributed.rpc.rpc_async` or :meth:`~torch.distributed.rpc.remote` are not idempotent and therefore cannot be retried. However, internal RRef control messages will be made idempotent and retryable.
• 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 living UserRRef instances and user code is not holding references to the OwnerRRef either. The tricky part is to determine if there are any living UserRRef 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 :meth:`~torch.distributed.rpc.rpc_sync`, :meth:`~torch.distributed.rpc.rpc_async`, or :meth:`~torch.distributed.rpc.remote` 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 any UserRRef 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 its OwnerRRef instance when it thinks there are no living UserRRef instances, 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.:

OwnerRRef -> A -> Y -> Z

If all of Z's messages, including the delete message, are processed by the owner before all messages from Y, the owner will learn Z's deletion before knowing Y. Nevertheless, this does not cause any problem. Because, at least one of Y's ancestors will be alive (in this case, 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 as the owner is A's parent.

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

OwnerRRef
    ^
    |
    A -> Y -> Z

If Z calls :meth:`~torch.distributed.rpc.RRef.to_here` on the UserRRef, the owner at least knows A when Z is deleted, because otherwise, :meth:`~torch.distributed.rpc.RRef.to_here` wouldn't finish. If Z does not call :meth:`~torch.distributed.rpc.RRef.to_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 UserRRef destructor. 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.

User Share RRef with Owner as Return Value

import torch
import torch.distributed.rpc as rpc

# on worker A
rref = rpc.remote('B', torch.add, args=(torch.ones(2), 1))
# say the rref has RRefId 100 and ForkId 1
rref.to_here()

In this case, the UserRRef is created on the user worker A, then it is passed to the owner worker B together with the remote message, and then B creates the OwnerRRef. The method :meth:`~torch.distributed.rpc.remote` returns immediately, meaning that the UserRRef can be forked/used before the owner knows about it.

On the owner, when receiving the :meth:`~torch.distributed.rpc.remote` call, it will create the OwnerRRef, and returns an ACK to acknowledge {100, 1} (RRefId, ForkId). Only after receiving this ACK, can A delete its UserRRef. This involves both G1 and G2. G1 is obvious. For G2, the OwnerRRef is a child of the UserRRef, and the UserRRef is not deleted until it receives the ACK from the owner.

The diagram above shows the message flow, where solid arrow contains user function and dashed arrow are builtin messages. Note that the first two messages from A to B (:meth:`~torch.distributed.rpc.remote` and :meth:`~torch.distributed.rpc.RRef.to_here`) may arrive at B in any order, but the final delete message will only be sent out when:

• B acknowledges UserRRef {100, 1} (G2), and
• Python GC agrees to delete the local UserRRef instance. This occurs when the RRef is no longer in scope and is eligible for garbage collection.

User Share RRef with Owner as Argument

import torch
import torch.distributed.rpc as rpc

# on worker A and worker B
def func(rref):
  pass

# on worker A
rref = rpc.remote('B', torch.add, args=(torch.ones(2), 1))
# say the rref has RRefId 100 and ForkId 1
rpc.rpc_async('B', func, args=(rref, ))

In this case, after creating the UserRRef on A, A uses it as an argument in a followup RPC call to B. A will keep UserRRef {100, 1} alive until it receives the acknowledge from B (G2, not the return value of the RPC call). This is necessary because A should not send out the delete message until all previous messages are received, otherwise, the OwnerRRef could be deleted before usage as we do not guarantee message delivery order. This is done by creating a child ForkId of RRef, holding them in a map until receives the owner confirms the child ForkId. The figure below shows the message flow.

Note that the UserRRef could be deleted on B before func finishes or even starts. However this is OK, as at the time B sends out ACK for the child ForkId, it already acquired the OwnerRRef instance, which would prevent it been deleted too soon.

Owner Share RRef with User

Owner to user is the simplest case, where the owner can update reference counting locally, and does not need any additional control message to notify others. Regarding G2, it is same as the parent receives the ACK from the owner immediately, as the parent is the owner.

import torch
import torch.distributed.rpc as RRef, rpc

# on worker B and worker C
def func(rref):
  pass

# on worker B, creating a local RRef
rref = RRef("data")
# say the rref has RRefId 100
dist.rpc_async('C', func, args=(rref, ))

The figure above shows the message flow. Note that when the OwnerRRef exits scope after the rpc_async call, it will not be deleted, because internally there is a map to hold it alive if there is any known forks, in which case is UserRRef {100, 1}. (G2)

User Share RRef with User

This is the most complicated case where caller user (parent UserRRef), callee user (child UserRRef), and the owner all need to get involved.

import torch
import torch.distributed.rpc as rpc

# on worker A and worker C
def func(rref):
  pass

# on worker A
rref = rpc.remote('B', torch.add, args=(torch.ones(2), 1))
# say the rref has RRefId 100 and ForkId 1
rpc.rpc_async('C', func, args=(rref, ))

When C receives the child UserRRef from A, it sends out a fork request to the owner B. Later, when the B confirms the UserRRef on C, C will perform two actions in parallel: 1) send out the child ACK to A ,and 2) run the user provided function. During this time, the parent (A) will hold its UserRRef {100, 1} alive to achieve G2.