NVIDIA DOCA Allreduce Application Guide

This guide provides a DOCA Allreduce collective operation implementation on top of NVIDIA® BlueField® DPU using UCX.

Introduction

Allreduce is a collective operation which allows collecting data from different processing units to combine them into a global result by a chosen operator. In turn, the result is distributed back to all processing units.

Allreduce operates in stages. Firstly, each participant scatters its vector. Secondly, each participant gathers the vectors of the other participants. Lastly, each participant performs their chosen operation between all the gathered vectors. Using a sequence of different allreduce operations with different participants, very complex computations can be spread among many computation units.

Allreduce is widely used by parallel applications in high-performance computing (HPC) related to scientific simulations and data analysis, including machine learning calculation and the training phase of neural networks in deep learning.

Due to the massive growth of deep learning models and the complexity of scientific simulation tasks that utilize a network, effective implementation of allreduce is essential for minimizing communication time.

This document describes how to implement allreduce using the UCX communication framework, which leverages NVIDIA® BlueField® DPU by providing low-latency and high-bandwidth utilization of its network engine.
This document describes the following types of allreduce:

Offloaded client – processes running on the host which only submit allreduce operation requests to a daemon running on the DPU. The daemon runs on the DPU and performs the allreduce algorithm on behalf of its on-host-clients (offloaded-client).
Non-offloaded client – processes running on the host which execute the allreduce algorithm by themselves

System Design

The application is designed to measure three metrics:

Communication time taken by offloaded and non-offloaded allreduce operations
Computation time taken by matrix multiplications which are done by clients until the allreduce operation is completed
The overlap of the two previous metrics. The percentage of the total runtime during which both the allreduce and the matrix multiplications were done in parallel.

system-design-diagram-version-1-modificationdate-1707753408817-api-v2.png

The allreduce implementation is divided into two different types of processes: clients and daemons. Clients are responsible for allocating vectors filled with data and initiating allreduce operations by sending a request with a vector to their daemon. Daemons are responsible for gathering vectors from all connected clients and daemons, applying a chosen operator on all received buffers, and then scattering the reduced result vector back to the clients.

Offloaded mode
Non-offloaded mode

Application Architecture

DOCA's allreduce implementation uses Unified Communication X (UCX) to support data exchange between endpoints. It utilizes UCX's sockaddr-based connection establishment and the UCX Active Messages (AM) API for communications.

Offloaded mode
Non-offloaded mode

Connections between processes are established by UCX using IP addresses and ports of peers.
Allreduce vectors are sent from clients to daemons in offloaded mode, or from clients to clients in non-offloaded mode.
Reduce operations on vectors are done using received vectors from other daemons in offloaded mode, or other clients in non-offloaded mode.
Vectors with allreduce results are received by clients from daemons in offloaded mode, or are already stored in clients after completing all exchanges in non-offloaded mode.
After completing all allreduce operations, connections between clients are destroyed.

DOCA Libraries

This application leverages the UCX framework DOCA driver.

Compiling the Application

Info

Please refer to the NVIDIA DOCA Installation Guide for Linux for details on how to install BlueField-related software.

The installation of DOCA's reference applications contains the sources of the applications, alongside the matching compilation instructions. This allows for compiling the applications "as-is" and provides the ability to modify the sources, then compile a new version of the application.

Tip

For more information about the applications as well as development and compilation tips, refer to the DOCA Applications page.

The sources of the application can be found under the application's directory: /opt/mellanox/doca/applications/allreduce/.

Compiling All Applications

All DOCA applications are defined under a single meson project. So, by default, the compilation includes all of them.

To build all the applications together, run:

Copy
Copied!

            
            cd /opt/mellanox/doca/applications/
meson /tmp/build 
ninja -C /tmp/build

Info

doca_allreduce is created under /tmp/build/allreduce/.

Compiling Only the Current Application

To build the allreduce application only:

Copy
Copied!

            
            cd /opt/mellanox/doca/applications/
meson /tmp/build -Denable_all_applications=false -Denable_allreduce=true
ninja -C /tmp/build

Info

doca_allreduce is created under /tmp/build/allreduce/.

Alternatively, the user can set the desired flags in the meson_options.txt file instead of providing them in the compilation command line:

Edit the following flags in /opt/mellanox/doca/applications/meson_options.txt:
- Set enable_all_applications to false
- Set enable_allreduce to true

Run the following compilation commands:

Copy
Copied!

            
            cd /opt/mellanox/doca/applications/
meson /tmp/build 
ninja -C /tmp/build

Info

doca_allreduce is created under /tmp/build/allreduce/.

Troubleshooting

Please refer to the NVIDIA DOCA Troubleshooting Guide for any issue encountered with the compilation of the application .

Running the Application

Application Execution

The allreduce application is provided in source form, hence a compilation is required before the application can be executed.

Application usage instructions:

Copy
Copied!

            
            Usage: doca_allreduce [DOCA Flags] [Program Flags]
 
DOCA Flags:
  -h, --help                        Print a help synopsis
  -v, --version                     Print program version information
  -l, --log-level                   Set the (numeric) log level for the program <10=DISABLE, 20=CRITICAL, 30=ERROR, 40=WARNING, 50=INFO, 60=DEBUG, 70=TRACE>
  --sdk-log-level                   Set the SDK (numeric) log level for the program <10=DISABLE, 20=CRITICAL, 30=ERROR, 40=WARNING, 50=INFO, 60=DEBUG, 70=TRACE>
  -j, --json <path>                 Parse all command flags from an input json file
 
Program Flags:
  -r, --role                        Run DOCA UCX allreduce process as: "client" or "daemon"
  -m, --mode <allreduce_mode>       Set allreduce mode: "offloaded", "non-offloaded" (valid for client only)
  -p, --port <port>                 Set default destination port of daemons/clients, used for IPs without a port (see '-a' flag)
  -t, --listen-port <listen_port>   Set listening port of daemon or client
  -c, --num-clients <num_clients>   Set the number of clients which participate in allreduce operations (valid for daemon only)
  -s, --size <size>                 Set size of vector to do allreduce for
  -d, --datatype <datatype>         Set datatype ("byte", "int", "float", "double") of vector elements to do allreduce for
  -o, --operation <operation>       Set operation ("sum", "prod") to do between allreduce vectors
  -b, --batch-size <batch_size>     Set the number of allreduce operations submitted simultaneously (used for handshakes by daemons)
  -i, --num-batches <num_batches>   Set the number of batches of allreduce operations (used for handshakes by daemons)
  -a, --address <ip_address>        Set comma-separated list of destination IPv4/IPv6 addresses and ports optionally (<ip_addr>:[<port>]) of daemons or clients

Info

This usage printout can be printed to the command line using the -h (or --help) options:

Copy
Copied!

            
            ./doca_allreduce -h

Info

For additional information, refer to section "Command Line Flags".

Configuration steps.

All daemons should be deployed before clients. Only after connecting to their peers are daemons able to handle clients.

UCX probes the system for any available net/IB devices and, by default, tries to create a multi-device connection. This means that if some network devices are available but provide an unreachable path from the daemon to the peer/client, UCX may still use that path. A common case is that a daemon tries to connect to a different BlueField using tmfifo_net0 which is connected to the host only. To fix this issue, follow these steps:

Use the UCX env variable UCX_NET_DEVICES to set usable devices. For example:

Copy
Copied!

            
            export UCX_NET_DEVICES=enp3s0f0s0,enp3s0f1s0
./doca_allreduce -r daemon -t 34001 -c 1 -s 100 -o sum -d float -b 16 -i 16

Or:

Copy
Copied!

            
            env UCX_NET_DEVICES=enp3s0f0s0,enp3s0f1s0 ./doca_allreduce -r daemon -t 34001 -c 1 -s 100 -o sum -d float -b 16 -i 16

Get the mlx device name and port of a SF to limit the UCX network interfaces and allow IB. For example:

Copy
Copied!

            
            BlueField> show_gids
DEV	PORT	INDEX	GID					IPv4  		VER	DEV
---	----	-----	---					------------        ---	---
mlx5_2	1	0	    fe80:0000:0000:0000:0052:72ff:fe63:1651		     v2	 enp3s0f0s0
mlx5_3	1	0	    fe80:0000:0000:0000:0032:6bff:fe13:f13a		     v2	 enp3s0f1s0
 
BlueField> UCX_NET_DEVICES=enp3s0f0s0,enp3s0f1s0,mlx5_2:1,mlx5_3:1 ./doca_allreduce -r daemon -t 34001 -c 1 -s 100 -o sum -d float -b 16 -i 16

CLI example for running the deamon on BlueField:
Copy

Copied!
```
            
            ./doca_allreduce -r daemon -t 34001 -c 2 -a 10.21.211.3:35001,10.21.211.4:36001 -s 65535 -o sum -d float -i 16 -b 128 
        
```
Notes:
- The flag -a is necessary for communicating with other daemons. In case of an offloaded client, the address must be that of the daemon which performs the allreduce operations for them. In case of a daemon or non-offloaded clients, the flag could be a single or multiple addresses of other daemons/non-offloaded clients which exchange their local allreduce results.
- The flag -c must be specified for daemon processes only. It indicates how many clients submit their allreduce operations to the daemon.
- The flags -s, -i, -b, and -d must be the same for all clients and daemons participating in the allreduce operation.
Note

The daemon listens to incoming connection requests on all available IPs, but the actual communication after the initial "UCX handshake" does not necessarily use the same device used for the connection establishment.

CLI example for running the client on the host:

Copy
Copied!

            
            ./doca_allreduce -r client -m non-offloaded -t 34001 -a 10.21.211.3:35001,10.21.211.4:36001 -s 65535 -i 16 -b 128 -o sum -d float
./doca_allreduce -r client -m offloaded -p 34001 -a 192.168.100.2 -s 65535 -i 16 -b 128 -o sum -d float

The application also supports a JSON-based deployment mode, in which all command-line arguments are provided through a JSON file:
Copy

Copied!
```
            
            ./doca_allreduce --json [json_file]
        
```
For example:
Copy

Copied!
```
            
            ./doca_allreduce --json ./allreduce_client_params.json
        
```
Note

Before execution, ensure that the used JSON file contains the correct configuration parameters, and especially the desired PCIe and network addresses required for the deployment.

Command Line Flags

Flag Type	Short Flag	Long Flag/JSON Key	Description
General flags	`h`	`help`	Print a help synopsis
	`v`	`version`	Print program version information
	`l`	`log-level`	Set the log level for the application: DISABLE=10 CRITICAL=20 ERROR=30 WARNING=40 INFO=50 DEBUG=60 TRACE=70 (requires compilation with `TRACE` log level support)
	N/A	`sdk-log-level`	Sets the log level for the program: DISABLE=10 CRITICAL=20 ERROR=30 WARNING=40 INFO=50 DEBUG=60 TRACE=70
	`j`	`json`	Parse all command flags from an input JSON file
Program flags	`r`	`role`	Run DOCA UCX allreduce process as either `client` or `daemon`
	`m`	`mode`	Set allreduce mode. Available types options: `offloaded` `non-offloaded` (valid for client only)
	`p`	`port`	Set default destination port of daemons/clients. Used for IPs without a port (see `-a` flag).
	`c`	`num-clients`	Set the number of clients which participate in allreduce operations Note: Valid for daemon only.
	`s`	`size`	Set size of vector to perform allreduce for
	`d`	`datatype`	Set datatype of vector elements to do allreduce for `byte` `int` `float` `double`
	`o`	`operation`	Set operation to perform between allreduce vectors
	`b`	`batch-size`	Set the number of allreduce operations submitted simultaneously. Used for handshakes by daemons.
	`i`	`num-batches`	Set the number of batches of allreduce operations. Used for handshakes by daemons.
	`t`	`listen-port`	Set listening port of daemon or client
	`a`	`address`	Set comma-separated list of destination IPv4/IPv6 address and ports optionally of daemons or clients. Format: `<ip_addr>:[<port>]`.

Info

Refer to DOCA Arg Parser for more information regarding the supported flags and execution modes.

Troubleshooting

Refer to the NVIDIA DOCA Troubleshooting Guide for any issue encountered with the installation or execution of the DOCA applications.

Running Application on NVIDIA Converged Accelerator

This section details the steps necessary to run DOCA Allreduce on NVIDIA converged accelerator.

Allreduce running on the converged accelerator has the same logic as described in previous sections except for the reducing of vectors. The reduce of incoming vectors is performed on the GPU side in batches that include the vectors from all peers or all clients. When the GPUDirect module is active, incoming vectors and outgoing vectors are received/sent directly to/from the GPU.

To make use of the GPU's capabilities, make sure to perform the following:

Refer to the NVIDIA DOCA Installation Guide for Linux for instructions on installing NVIDIA driver for CUDA and a CUDA-repo on your setup.
Create the sub-functions and configure the OVS according to NVIDIA BlueField DPU Scalable Function User Guide.

Compiling and Running Application

To build and run the application:

Setup CUDA paths:

Copy
Copied!

            
            export CPATH=/usr/local/cuda/targets/sbsa-linux/include:$CPATH
export LD_LIBRARY_PATH=/usr/local/nvidia/lib:/usr/local/nvidia/lib64:$LD_LIBRARY_PATH
export PATH=/usr/local/nvidia/bin:/usr/local/cuda/bin:$PATH

Reinstall UCX with CUDA support. Follow the UCX installation procedure with an additional flag, --with-cuda=/usr/local/cuda/, passed to configure-release:
Copy

Copied!
```
            
            dpu# ./contrib/configure-release --with-cuda=/usr/local/cuda/
        
```
To build the application with GPU support:
1. Set the enable_gpu_support flag to true in /opt/mellanox/doca/applications/meson_option.txt.
2. Compile the application sources. Run:
  Copy
  
  Copied!
```
            
            cd /opt/mellanox/doca/applications/
meson /tmp/build
ninja -C /tmp/build
        
```
  doca_allreduce_gpu is created under /tmp/build/allreduce/ alongside the regular doca_allreduce binary that is compiled without the GPU support.
To run the application with GPU support, follow the same steps as described in section "Running the Application".

Application Code Flow

Parse application argument.

Initialize arg parser resources and register DOCA general parameters.

Copy
Copied!

            
            doca_argp_init();

Register UCX allreduce application parameters.

Copy
Copied!

            
            register_allreduce_params();

Parse all registered parameters.

Copy
Copied!

            
            doca_argp_start();

UCX initialization.

Initialize hash table of connections.

Copy
Copied!

            
            allreduce_ucx_init();

Create UCP context.

Copy
Copied!

            
            ucp_init();

Create UCP worker.

Copy
Copied!

            
            ucp_worker_create();

Set AM handler for receiving connection check packets.

Copy
Copied!

            
            ucp_worker_set_am_recv_handler();

Initialization of the allreduce connectivity.
Copy

Copied!
```
            
            communication_init();
        
```
1. Initialize hash table of allreduce super requests.
2. Set "receive callback" for handshake messages.
3. If daemon or non-offloaded client:
  1. Set AM handler for receiving allreduce requests from clients.
    Copy
    
    Copied!
    
    allreduce_ucx_am_set_recv_handler();
  2. Initialize UCX listening function. This creates a UCP listener.
    Copy
    
    Copied!
    
    allreduce_ucx_listen();
4. Initialize all connections.
  Copy
  
  Copied!
```
            
            connections_init();
        
```
  1. Go over all destination addresses and connect to each peer.
  2. Repeat until a successful send occurs (to check connectivity).
    Copy
    
    Copied!
    
    ucp_am_send_nbx(); allreduce_ucx_request_wait();
  3. Insert the connection to the hash table of connections.
    Copy
    
    Copied!
    
    allreduce_outgoing_handshake();
5. Scatter handshake message to peers/daemon to make sure they all have the same -s, -i, -b, and -d flags.

Daemon: Start UCX progress.

Copy
Copied!

            
            daemon_run();

Set AM handler to receive allreduce requests from clients.

Copy
Copied!

            
            allreduce_ucx_am_set_recv_handler();

Perform UCP worker progress.

Copy
Copied!

            
            while (running)
    allreduce_ucx_progress();

Callbacks are invoked by incoming/outgoing messages by calling allreduce_ucx_progress.

Client:
Copy

Copied!
```
            
            client_run();
        
```
1. Allocate buffers to store allreduce initial data and results.
  Copy
  
  Copied!
```
            
            allreduce_vectors_init();
        
```
2. Set an AM handler for receiving allreduce results.
  Copy
  
  Copied!
```
            
            allreduce_ucx_am_set_recv_handler();
        
```
3. Perform allreduce barrier. Check that all daemons and clients are active.
  Copy
  
  Copied!
```
            
            allreduce_barrier();
        
```
  1. Submit a batch of allreduce operations with 0 byte.
  2. Wait for completions.
4. Reset metrics and vectors.
  Copy
  
  Copied!
```
            
            allreduce_metrics_init();
        
```
  1. Submit some batches and calculate estimated network time.
  2. Allocate matrices to multiply.
  3. Estimate how many matrix multiplications could have been performed instead of networking (same time window).
  4. Calculate the actual computation time of these matrix multiplications.
5. Reset vectors.
6. Submit a batch of allreduce operations to daemon/peer (depends on mode).
7. Perform matrix multiplications during a time period which is approximately equal to doing a single batch of allreduce operations and calculate the actual time cost.
8. Wait for the allreduce operation to complete and calculate time cost.
9. Update metrics.
  Copy
  
  Copied!
```
            
            Do num-batches (flag) times:
    allreduce_vectors_reset();
    allreduce_batch_submit();
    cpu_exploit();
    allreduce_batch_wait();
    allreduce_metrics_calculate();
        
```
10. Print summary of allreduce benchmarking.
  Copy
  
  Copied!
```
            
            allreduce_metrics_print();
        
```

Arg parser destroy.

Copy
Copied!

            
            doca_argp_destroy();

Communication destroy.
1. Clean up connections.
  Copy
  
  Copied!
```
            
            allreduce_ucx_disconnect();
        
```
  1. Remove the connection from the hash table of the connections.
  2. Close inner UCP endpoint.
    Copy
    
    Copied!
    
    ucp_ep_close_nbx();
  3. Wait for the completion of the UCP endpoint closure.
  4. Destroy connection.
  5. Free connections array.
2. Destroy the hash table of the allreduce super requests.

Destroy UCX context.

Destroy the hash table of the connections.

Copy
Copied!

            
            g_hash_table_destroy();

If the UCP listener was created, destroy it.

Copy
Copied!

            
            ucp_listener_destroy();

Destroy UCP worker.

Copy
Copied!

            
            ucp_worker_destroy();

Destroy UCP context.

Copy
Copied!

            
            ucp_cleanup();

References

/opt/mellanox/doca/applications/allreduce/
/opt/mellanox/doca/applications/allreduce/allreduce_client_params.json
/opt/mellanox/doca/applications/allreduce/allreduce_daemon_params.json

On This Page