Configuration#

UCX/UCXX can be configured with a wide variety of options and optimizations including: transport, caching, etc. Users can configure UCX/UCXX either with environment variables or programmatically during initialization. Below we demonstrate setting UCX_MEMTYPE_CACHE to n and checking the configuration:

import ucxx
options = {"PROTO_INFO": "y"}
ucxx.init(options)
assert ucxx.get_config()['PROTO_INFO'] == 'y'

Note

When programmatically configuring UCXX, the UCX prefix is not used.

For novice users we recommend using UCXX defaults, see the next section for details.

UCXX vs UCX Defaults#

UCXX redefines some of the UCX defaults for a variety of reasons, including better performance for the more common Python use cases, or to work around known limitations or bugs of UCX. To verify UCX default configurations, for the currently installed UCX version please run the command-line tool ucx_info -f.

Below is a list of the UCXX redefined default values, and what conditions are required for them to apply.

Apply to all UCX versions:

UCX_MEMTYPE_CACHE=n
UCX_RNDV_THRESH=8192
UCX_FRAG_MEM_TYPE=cuda
UCX_MAX_RNDV_RAILS=1

Apply to UCX < 1.18.0, newer versions rely on UCX defaults:

UCX_PROTO_ENABLE=n

Apply to devices with a BAR1 size equal to the device size, provided it’s not in MIG mode:

UCX_CUDA_COPY_MAX_REG_RATIO=1.0

Please note that UCX_CUDA_COPY_MAX_REG_RATIO=1.0 is only set provided at least one GPU is present with a BAR1 size smaller than its total memory (e.g., NVIDIA T4).

UCX Environment Variables in UCXX#

In this section we go over a brief overview of some of the more relevant variables for current UCXX usage, along with some comments on their uses and limitations. To see a complete list of UCX environment variables, their descriptions and default values, please run the command-line tool ucx_info -f.

UCP CONTEXT CONFIGURATION#

Configuration variables applying to the UCP context.

UCX_PROTO_ENABLE#

Values: y, n

Enable the new protocol selection logic, also known as “protov2”. Its default has been changed to y starting with UCX 1.16.0.

The new protocol solves various limitations from the original “protov1” including, for example, invalid choice of transport in systems with hybrid interconnectivity, such as a DGX-1 where only a subset of GPU pairs are interconnected via NVLink.

DEBUG#

Debug variables for both UCX and UCXX can be set

UCXPY_LOG_LEVEL/UCX_LOG_LEVEL#

Values: DEBUG, TRACE

If UCX has been built with debug mode enabled

MEMORY#

UCX_MEMTYPE_CACHE#

This is a UCX Memory optimization which toggles whether UCX library intercepts cu*alloc* calls. UCXX defaults this value to n. There known issues when using this feature.

Values: n/y

UCX_CUDA_IPC_CACHE#

This is a UCX CUDA Memory optimization which enables/disables a remote endpoint IPC memhandle mapping cache. UCX/UCXX defaults this value to y

Values: n/y

UCX_MEMTYPE_REG_WHOLE_ALLOC_TYPES#

By defining UCX_MEMTYPE_REG_WHOLE_ALLOC_TYPES=cuda (default in UCX >= 1.12.0), UCX enables registration cache based on a buffer’s base address, thus preventing multiple time-consuming registrations for the same buffer. This is particularly useful when using a CUDA memory pool, thus requiring a single registration between two ends for the entire pool, providing considerable performance gains, especially when using InfiniBand.

TRANSPORTS#

UCX_MAX_RNDV_RAILS#

Limiting the number of rails (network devices) to 1 allows UCX to use only the closest device according to NUMA locality and system topology. Particularly useful with InfiniBand and CUDA GPUs, ensuring all transfers from/to the GPU will use the closest InfiniBand device and thus implicitly enable GPUDirectRDMA.

Note

On CPU-only systems, better network bandwidth performance with infiniband transports may be achieved by letting UCX use more than a single network device. This can be achieved by explicitly setting UCX_MAX_RNDV_RAILS to 2 or higher.

Values: Int (UCXX default: 1)

UCX_RNDV_THRESH#

This is a configurable parameter used by UCX to help determine which transport method should be used. For example, on machines with multiple GPUs, and with NVLink enabled, UCX can deliver messages either through TCP or NVLink. Sending GPU buffers over TCP is costly as it triggers a device-to-host on the sender side, and then host-to-device transfer on the receiver side – we want to avoid these kinds of transfers when NVLink is available. If a buffer is below the threshold, Rendezvous-Protocol is triggered and for UCXX users, this will typically mean messages will be delivered through TCP. Depending on the application, messages can be quite small, therefore, we recommend setting a small value if the application uses NVLink or InfiniBand: UCX_RNDV_THRESH=8192

Values: Int (UCXX default: 8192)

UCX_RNDV_SCHEME#

Communication scheme in RNDV protocol

Values:

  • put_zcopy

  • get_zcopy

  • auto (default)

UCX_TCP_RX_SEG_SIZE#

Size of send copy-out buffer when receiving. This environment variable controls the size of the buffer on the host when receiving data over TCP.

UCX_TCP_TX_SEG_SIZE#

Size of send copy-out buffer when transmitting. This environment variable controls the size of the buffer on the host when sending data over TCP.

Note

Users should take care to properly tune UCX_TCP_{RX/TX}_SEG_SIZE parameters when mixing TCP with other transports methods as well as when using TCP over UCX in isolation. These variables will impact CUDA transfers when no NVLink or InfiniBand is available between UCXX processes. These parameters will cause the HostToDevice and DeviceToHost copies of buffers to be broken down in several chunks when the size of a buffer exceeds the size defined by these two variables. If an application is expected to transfer very large buffers, increasing such values may improve overall performance.

UCX_TLS#

Transport Methods (Simplified):

  • all -> use all the available transports

  • rc -> InfiniBand (ibv_post_send, ibv_post_recv, ibv_poll_cq) uses rc_v and rc_x (preferably if available)

  • cuda_copy -> cuMemHostRegister, cuMemcpyAsync

  • cuda_ipc -> CUDA Interprocess Communication (cuIpcCloseMemHandle, cuIpcOpenMemHandle, cuMemcpyAsync)

  • sm/shm -> all shared memory transports (mm, cma, knem)

  • mm -> shared memory transports - only memory mappers

  • ugni -> ugni_smsg and ugni_rdma (uses ugni_udt for bootstrap)

  • ib -> all infiniband transports (rc/rc_mlx5, ud/ud_mlx5, dc_mlx5)

  • rc_v -> rc verbs (uses ud for bootstrap)

  • rc_x -> rc with accelerated verbs (uses ud_mlx5 for bootstrap)

  • ud_v -> ud verbs

  • ud_x -> ud with accelerated verbs

  • ud -> ud_v and ud_x (preferably if available)

  • dc/dc_x -> dc with accelerated verbs

  • tcp -> sockets over TCP/IP

  • cuda -> CUDA (NVIDIA GPU) memory support

  • rocm -> ROCm (AMD GPU) memory support

SOCKADDR_TLS_PRIORITY#

Priority of sockaddr transports

InfiniBand Device#

Select InfiniBand Device

UCX_NET_DEVICES#

It’s recommended to not define this variable and instead let UCX determine the closest InfiniBand device, note that this required the CUDA context to be created before UCX/UCXX are initialized.

Typically these will be the InfiniBand device corresponding to a particular set of GPUs. Values:

  • Check ucx-info -d for acomplete list

To find more information on the topology of InfiniBand-GPU pairing run the following:

nvidia-smi topo -m

Example Configs#