Ray Collective Communication Lib#

The Ray collective communication library (ray.util.collective) offers a set of native collective primitives for communication between distributed CPUs or GPUs.

Ray collective communication library

enables 10x more efficient out-of-band collective communication between Ray actor and task processes,
operates on both distributed CPUs and GPUs,
uses NCCL and GLOO as the optional high-performance communication backends,
is suitable for distributed ML programs on Ray.

Collective Primitives Support Matrix#

See below the current support matrix for all collective calls with different backends.

Backend	gloo		nccl
Device	CPU	GPU	CPU	GPU
send	✔	✘	✘	✔
recv	✔	✘	✘	✔
broadcast	✔	✘	✘	✔
allreduce	✔	✘	✘	✔
reduce	✔	✘	✘	✔
allgather	✔	✘	✘	✔
gather	✘	✘	✘	✘
scatter	✘	✘	✘	✘
reduce_scatter	✔	✘	✘	✔
all-to-all	✘	✘	✘	✘
barrier	✔	✘	✘	✔

Supported Tensor Types#

torch.Tensor
numpy.ndarray
cupy.ndarray

Usage#

Installation and Importing#

Ray collective library is bundled with the released Ray wheel. Besides Ray, users need to install either pygloo or cupy in order to use collective communication with the GLOO and NCCL backend, respectively.

pip install pygloo
pip install cupy-cudaxxx # replace xxx with the right cuda version in your environment

To use these APIs, import the collective package in your actor/task or driver code via:

import ray.util.collective as col

Initialization#

Collective functions operate on collective groups. A collective group contains a number of processes (in Ray, they are usually Ray-managed actors or tasks) that will together enter the collective function calls. Before making collective calls, users need to declare a set of actors/tasks, statically, as a collective group.

Below is an example code snippet that uses the two APIs init_collective_group() and declare_collective_group() to initialize collective groups among a few remote actors. Refer to APIs for the detailed descriptions of the two APIs.

import ray
import ray.util.collective as collective

import cupy as cp


@ray.remote(num_gpus=1)
class Worker:
   def __init__(self):
       self.send = cp.ones((4, ), dtype=cp.float32)
       self.recv = cp.zeros((4, ), dtype=cp.float32)

   def setup(self, world_size, rank):
       collective.init_collective_group(world_size, rank, "nccl", "default")
       return True

   def compute(self):
       collective.allreduce(self.send, "default")
       return self.send

   def destroy(self):
       collective.destroy_group()

# imperative
num_workers = 2
workers = []
init_rets = []
for i in range(num_workers):
   w = Worker.remote()
   workers.append(w)
   init_rets.append(w.setup.remote(num_workers, i))
_ = ray.get(init_rets)
results = ray.get([w.compute.remote() for w in workers])


# declarative
for i in range(num_workers):
   w = Worker.remote()
   workers.append(w)
_options = {
   "group_name": "177",
   "world_size": 2,
   "ranks": [0, 1],
   "backend": "nccl"
}
collective.declare_collective_group(workers, **_options)
results = ray.get([w.compute.remote() for w in workers])

Note that for the same set of actors/task processes, multiple collective groups can be constructed, with group_name as their unique identifier. This enables to specify complex communication patterns between different (sub)set of processes.

Collective Communication#

Check the support matrix for the current status of supported collective calls and backends.

Note that the current set of collective communication API are imperative, and exhibit the following behaviours:

All the collective APIs are synchronous blocking calls
Since each API only specifies a part of the collective communication, the API is expected to be called by each participating process of the (pre-declared) collective group. Upon all the processes have made the call and rendezvous with each other, the collective communication happens and proceeds.
The APIs are imperative and the communication happends out-of-band — they need to be used inside the collective process (actor/task) code.

An example of using ray.util.collective.allreduce is below:

import ray
import cupy
import ray.util.collective as col


@ray.remote(num_gpus=1)
class Worker:
    def __init__(self):
        self.buffer = cupy.ones((10,), dtype=cupy.float32)

    def compute(self):
        col.allreduce(self.buffer, "default")
        return self.buffer

# Create two actors A and B and create a collective group following the previous example...
A = Worker.remote()
B = Worker.remote()
# Invoke allreduce remotely
ray.get([A.compute.remote(), B.compute.remote()])

Point-to-point Communication#

ray.util.collective also supports P2P send/recv communication between processes.

The send/recv exhibits the same behavior with the collective functions: they are synchronous blocking calls – a pair of send and recv must be called together on paired processes in order to specify the entire communication, and must successfully rendezvous with each other to proceed. See the code example below:

import ray
import cupy
import ray.util.collective as col


@ray.remote(num_gpus=1)
class Worker:
    def __init__(self):
        self.buffer = cupy.ones((10,), dtype=cupy.float32)

    def get_buffer(self):
        return self.buffer

    def do_send(self, target_rank=0):
        # this call is blocking
        col.send(target_rank)

    def do_recv(self, src_rank=0):
        # this call is blocking
        col.recv(src_rank)

    def do_allreduce(self):
        # this call is blocking as well
        col.allreduce(self.buffer)
        return self.buffer

# Create two actors
A = Worker.remote()
B = Worker.remote()

# Put A and B in a collective group
col.declare_collective_group([A, B], options={rank=[0, 1], ...})

# let A to send a message to B; a send/recv has to be specified once at each worker
ray.get([A.do_send.remote(target_rank=1), B.do_recv.remote(src_rank=0)])

# An anti-pattern: the following code will hang, because it doesn't instantiate the recv side call
ray.get([A.do_send.remote(target_rank=1)])

Single-GPU and Multi-GPU Collective Primitives#

In many cluster setups, a machine usually has more than 1 GPU; effectively leveraging the GPU-GPU bandwidth, such as NVLINK, can significantly improve communication performance.

ray.util.collective supports multi-GPU collective calls, in which case, a process (actor/tasks) manages more than 1 GPU (e.g., via ray.remote(num_gpus=4)). Using these multi-GPU collective functions are normally more performance-advantageous than using single-GPU collective API and spawning the number of processes equal to the number of GPUs. See the API references for the signatures of multi-GPU collective APIs.

Also of note that all multi-GPU APIs are with the following restrictions:

Only NCCL backend is supported.
Collective processes that make multi-GPU collective or P2P calls need to own the same number of GPU devices.
The input to multi-GPU collective functions are normally a list of tensors, each located on a different GPU device owned by the caller process.

An example code utilizing the multi-GPU collective APIs is provided below:

import ray
import ray.util.collective as collective

import cupy as cp
from cupy.cuda import Device


@ray.remote(num_gpus=2)
class Worker:
   def __init__(self):
       with Device(0):
           self.send1 = cp.ones((4, ), dtype=cp.float32)
       with Device(1):
           self.send2 = cp.ones((4, ), dtype=cp.float32) * 2
       with Device(0):
           self.recv1 = cp.ones((4, ), dtype=cp.float32)
       with Device(1):
           self.recv2 = cp.ones((4, ), dtype=cp.float32) * 2

   def setup(self, world_size, rank):
       self.rank = rank
       collective.init_collective_group(world_size, rank, "nccl", "177")
       return True

   def allreduce_call(self):
       collective.allreduce_multigpu([self.send1, self.send2], "177")
       return [self.send1, self.send2]

   def p2p_call(self):
       if self.rank == 0:
          collective.send_multigpu(self.send1 * 2, 1, 1, "8")
       else:
          collective.recv_multigpu(self.recv2, 0, 0, "8")
       return self.recv2

# Note that the world size is 2 but there are 4 GPUs.
num_workers = 2
workers = []
init_rets = []
for i in range(num_workers):
   w = Worker.remote()
   workers.append(w)
   init_rets.append(w.setup.remote(num_workers, i))
a = ray.get(init_rets)
results = ray.get([w.allreduce_call.remote() for w in workers])
results = ray.get([w.p2p_call.remote() for w in workers])

More Resources#

The following links provide helpful resources on how to efficiently leverage the ray.util.collective library.

More running examples under ray.util.collective.examples.
Scaling up the Spacy Name Entity Recognition (NER) pipeline using Ray collective library.
Implementing the AllReduce strategy for data-parallel distributed ML training.

API References#

APIs exposed under the namespace ray.util.collective.

class ray.util.collective.collective.GroupManager[source]#

Use this class to manage the collective groups we created so far.

Each process will have an instance of GroupManager. Each process could belong to multiple collective groups. The membership information and other metadata are stored in the global _group_mgr object.

create_collective_group(backend, world_size, rank, group_name)[source]#

The entry to create new collective groups in the manager.

Put the registration and the group information into the manager metadata as well.

get_group_by_name(group_name)[source]#: Get the collective group handle by its name.

destroy_collective_group(group_name)[source]#: Group destructor.

ray.util.collective.collective.is_group_initialized(group_name)[source]#: Check if the group is initialized in this process by the group name.

ray.util.collective.collective.init_collective_group(world_size: int, rank: int, backend='nccl', group_name: str = 'default')[source]#

Initialize a collective group inside an actor process.

Parameters:

world_size – the total number of processes in the group.
rank – the rank of the current process.
backend – the CCL backend to use, NCCL or GLOO.
group_name – the name of the collective group.

Returns:

None

ray.util.collective.collective.create_collective_group(actors, world_size: int, ranks: List[int], backend='nccl', group_name: str = 'default')[source]#

Declare a list of actors as a collective group.

Note: This function should be called in a driver process.

Parameters:

actors – a list of actors to be set in a collective group.
world_size – the total number of processes in the group.
ranks (List[int]) – the rank of each actor.
backend – the CCL backend to use, NCCL or GLOO.
group_name – the name of the collective group.

Returns:

None

ray.util.collective.collective.destroy_collective_group(group_name: str = 'default') → None[source]#: Destroy a collective group given its group name.

ray.util.collective.collective.get_rank(group_name: str = 'default') → int[source]#

Return the rank of this process in the given group.

Parameters:: group_name – the name of the group to query
Returns:: the rank of this process in the named group, -1 if the group does not exist or the process does not belong to the group.

ray.util.collective.collective.get_collective_group_size(group_name: str = 'default') → int[source]#

Return the size of the collective group with the given name.

Parameters:

group_name – the name of the group to query

Returns:

The world size of the collective group, -1 if the group does: not exist or the process does not belong to the group.

ray.util.collective.collective.allreduce(tensor, group_name: str = 'default', op=ReduceOp.SUM)[source]#

Collective allreduce the tensor across the group.

Parameters:

tensor – the tensor to be all-reduced on this process.
group_name – the collective group name to perform allreduce.
op – The reduce operation.

Returns:

None

ray.util.collective.collective.allreduce_multigpu(tensor_list: list, group_name: str = 'default', op=ReduceOp.SUM)[source]#

Collective allreduce a list of tensors across the group.

Parameters:

tensor_list (List[tensor]) – list of tensors to be allreduced, each on a GPU.
group_name – the collective group name to perform allreduce.

Returns:

None

ray.util.collective.collective.barrier(group_name: str = 'default')[source]#

Barrier all processes in the collective group.

Parameters:: group_name – the name of the group to barrier.
Returns:: None

ray.util.collective.collective.reduce(tensor, dst_rank: int = 0, group_name: str = 'default', op=ReduceOp.SUM)[source]#

Reduce the tensor across the group to the destination rank.

Parameters:

tensor – the tensor to be reduced on this process.
dst_rank – the rank of the destination process.
group_name – the collective group name to perform reduce.
op – The reduce operation.

Returns:

None

ray.util.collective.collective.reduce_multigpu(tensor_list: list, dst_rank: int = 0, dst_tensor: int = 0, group_name: str = 'default', op=ReduceOp.SUM)[source]#

Reduce the tensor across the group to the destination rank and destination tensor.

Parameters:

tensor_list – the list of tensors to be reduced on this process; each tensor located on a GPU.
dst_rank – the rank of the destination process.
dst_tensor – the index of GPU at the destination.
group_name – the collective group name to perform reduce.
op – The reduce operation.

Returns:

None

ray.util.collective.collective.broadcast(tensor, src_rank: int = 0, group_name: str = 'default')[source]#

Broadcast the tensor from a source process to all others.

Parameters:

tensor – the tensor to be broadcasted (src) or received (destination).
src_rank – the rank of the source process.
group_name – the collective group name to perform broadcast.

Returns:

None

ray.util.collective.collective.broadcast_multigpu(tensor_list, src_rank: int = 0, src_tensor: int = 0, group_name: str = 'default')[source]#

Broadcast the tensor from a source GPU to all other GPUs.

Parameters:

tensor_list – the tensors to broadcast (src) or receive (dst).
src_rank – the rank of the source process.
src_tensor – the index of the source GPU on the source process.
group_name – the collective group name to perform broadcast.

Returns:

None

ray.util.collective.collective.allgather(tensor_list: list, tensor, group_name: str = 'default')[source]#

Allgather tensors from each process of the group into a list.

Parameters:

tensor_list – the results, stored as a list of tensors.
tensor – the tensor (to be gathered) in the current process
group_name – the name of the collective group.

Returns:

None

ray.util.collective.collective.allgather_multigpu(output_tensor_lists: list, input_tensor_list: list, group_name: str = 'default')[source]#

Allgather tensors from each gpus of the group into lists.

Parameters:

output_tensor_lists (List[List[tensor]]) – gathered results, with shape must be num_gpus * world_size * shape(tensor).
input_tensor_list – (List[tensor]): a list of tensors, with shape num_gpus * shape(tensor).
group_name – the name of the collective group.

Returns:

None

ray.util.collective.collective.reducescatter(tensor, tensor_list: list, group_name: str = 'default', op=ReduceOp.SUM)[source]#

Reducescatter a list of tensors across the group.

Reduce the list of the tensors across each process in the group, then scatter the reduced list of tensors – one tensor for each process.

Parameters:

tensor – the resulted tensor on this process.
tensor_list – The list of tensors to be reduced and scattered.
group_name – the name of the collective group.
op – The reduce operation.

Returns:

None

ray.util.collective.collective.reducescatter_multigpu(output_tensor_list, input_tensor_lists, group_name: str = 'default', op=ReduceOp.SUM)[source]#

Reducescatter a list of tensors across all GPUs.

Parameters:

output_tensor_list – the resulted list of tensors, with shape: num_gpus * shape(tensor).
input_tensor_lists – the original tensors, with shape: num_gpus * world_size * shape(tensor).
group_name – the name of the collective group.
op – The reduce operation.

Returns:

None.

ray.util.collective.collective.send(tensor, dst_rank: int, group_name: str = 'default')[source]#

Send a tensor to a remote process synchronously.

Parameters:

tensor – the tensor to send.
dst_rank – the rank of the destination process.
group_name – the name of the collective group.

Returns:

None

ray.util.collective.collective.send_multigpu(tensor, dst_rank: int, dst_gpu_index: int, group_name: str = 'default', n_elements: int = 0)[source]#

Send a tensor to a remote GPU synchronously.

The function asssume each process owns >1 GPUs, and the sender process and receiver process has equal nubmer of GPUs.

Parameters:

tensor – the tensor to send, located on a GPU.
dst_rank – the rank of the destination process.
dst_gpu_index – the destination gpu index.
group_name – the name of the collective group.
n_elements – if specified, send the next n elements from the starting address of tensor.

Returns:

None

ray.util.collective.collective.recv(tensor, src_rank: int, group_name: str = 'default')[source]#

Receive a tensor from a remote process synchronously.

Parameters:

tensor – the received tensor.
src_rank – the rank of the source process.
group_name – the name of the collective group.

Returns:

None

ray.util.collective.collective.recv_multigpu(tensor, src_rank: int, src_gpu_index: int, group_name: str = 'default', n_elements: int = 0)[source]#

Receive a tensor from a remote GPU synchronously.

The function asssume each process owns >1 GPUs, and the sender process and receiver process has equal nubmer of GPUs.

Parameters:

tensor – the received tensor, located on a GPU.
src_rank – the rank of the source process.
src_gpu_index (int)
group_name – the name of the collective group.

Returns:

None

ray.util.collective.collective.synchronize(gpu_id: int)[source]#

Synchronize the current process to a give device.

Parameters:: gpu_id – the GPU device id to synchronize.
Returns:: None