Giters

acoustic-warfare / data-transfer

Data-transfer between COTS devices using GPUDirect RDMA over UCX

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

cythonnvidiaucx-pyccudagpudirect-rdma

data-transfer

A COTS GPUDirect RDMA data transer demo between two nodes (Back-to-Back) using Mellanox ConnectX-6 Dx and Nvidia Quadro M4000 GPU's over UCX.

Transfer Demo

Algorithm

Algorithm

Installation

Clone the project:

git clone https://github.com/acoustic-warfare/data-transfer.git

Move to the project folder:

cd data-transfer

Requirements

Hardware

1. NICs

  • ConnectX-6 Lx
  • ConnectX-6 Dx (Ours)
  • ConnectX-6
  • ConnectX-5
  • ConnectX-4 Lx

2. GPUs

Nvidia:

  • Tesla™ (any)
  • Quadro™ K-Series
  • Quadro™ P-Series (Ours)

Software

  • NIC In order to use the network interface cards, the package NVIDIA Firmware Tools: mft must be installed for firmware management together with correct drivers for Linux MLNX_OFED.

  • GPU In order to utilize GPUDirect RDMA, the package nvidia-peer-mem must be installed. MLNX_OFED 5.1

  • Data Transfer To be able to control the host channel adapter (HCA), the HPC networking library ucx is required with support for GPUDirect RDMA.

Installation

NVIDIA Drivers

It is preffered to install display drivers using the distribution's native package management tool, i.e apt. If not installed already, NVIDIA display drivers can be installed from NVIDIA Download Center.

1. CUDA Runtime and Toolkit

To install CUDA Toolkit CUDA 11.7 from Nvidia on Ubuntu 20.04:

wget https://developer.download.nvidia.com/compute/cuda/repos/ubuntu2004/x86_64/cuda-ubuntu2004.pin
sudo mv cuda-ubuntu2004.pin /etc/apt/preferences.d/cuda-repository-pin-600
wget https://developer.download.nvidia.com/compute/cuda/11.7.0/local_installers/cuda-repo-ubuntu2004-11-7-local_11.7.0-515.43.04-1_amd64.deb

sudo dpkg -i cuda-repo-ubuntu2004-11-7-local_11.7.0-515.43.04-1_amd64.deb
sudo cp /var/cuda-repo-ubuntu2004-11-7-local/cuda-*-keyring.gpg /usr/share/keyrings/

sudo apt-get update
sudo apt-get -y install cuda

2. MLNX_OFED NIC Firmware (v2.1-x.x.x or later)

To install Linux drivers for ethernet and infiniband adapters, MLNX_OFED. See Download Center or for Ubuntu 20.04:

wget http://www.mellanox.com/downloads/ofed/MLNX_OFED-5.6-2.0.9.0/MLNX_OFED_LINUX-5.6-2.0.9.0-ubuntu20.04-x86_64.tgz
tar -xvf MLNX_OFED_LINUX*

cd MLNX_OFED_LINUX*
sudo ./mlnxofedinstall --upstream-libs --dpdk --force

sudo /etc/init.d/openibd restart

3. GPUDirect RDMA

To install nvidia-peer-mem on Ubuntu 20.04:

git clone https://github.com/Mellanox/nv_peer_memory.git
cd nv_peer_memory
./build_module.sh

cd /tmp
tar xzf /tmp/nvidia-peer-memory_*
cd nvidia-peer-memory-*
dpkg-buildpackage -us -uc
sudo dpkg -i /tmp/nvidia-peer-memory_*.deb

sudo service nv_peer_mem restart

3.1 GDRCopy (OPTIONAL)

To install gdrcopy

git clone https://github.com/NVIDIA/gdrcopy.git
cd gdrcopy
sudo apt install check libsubunit0 libsubunit-dev
mkdir final
make prefix=final CUDA="$CUDA_HOME" all install

3.2 Performance Benchmark (OPTIONAL)

sudo apt update -y
sudo apt install -y libpci-dev libibumad

git clone https://github.com/linux-rdma/perftest.git
cd perftest
./autogen.sh && ./configure CUDA_H_PATH=/usr/local/cuda/include/cuda.h && make -j
sudo make install

4. UCX

To install ucx on Ubuntu 20.04, python bindings are required together with python3+ packages. To maintain a working environment, installing conda is highly recomended.

4.1 Conda

wget https://repo.anaconda.com/miniconda/Miniconda3-latest-Linux-x86.sh

bash Miniconda3-latest-Linux-x86.sh

Disable auto-init

conda config --set auto_activate_base false

Recreate an environment inside data-transfer

conda env create -f environment.yml

Activate and enter the newly created environment

conda activate data-transfer
...
(data-transfer) $

4.2 Development packages for the program

Ensure that the latest updates are installed.

sudo apt update -y

Install development packages in Ubuntu 20.04. libnuma-dev, cython3.

sudo apt install -y libnuma-dev cython3

4.3 Install dependencies

Some python modules require more dependencies to run.

pip install pynvml cpython

To build the project requires having the gcc compiler.

sudo apt install gcc cmake

Setup

The adapters need assigned IP addresses, we used the GNOME Network Manager GUI and assigned the IPv4 addresses 10.0.0.x format with netmask 255.255.255.0

To run the demo, make sure both network adapters are connected and working properly by performing a ping. It is important to specify the correct interface when pinging:

ping -I <LOCAL_INTERFACE> <REMOTE_ADDRESS>

Eg:

(data-transfer) scarecrow@node1:~/data-transfer$ ping -I ens4f0np0 10.0.0.4
PING 10.0.0.4 (10.0.0.4) from 10.0.0.3 ens4f0np0: 56(84) bytes of data.
64 bytes from 10.0.0.4: icmp_seq=1 ttl=64 time=0.110 ms
64 bytes from 10.0.0.4: icmp_seq=2 ttl=64 time=0.112 ms
64 bytes from 10.0.0.4: icmp_seq=3 ttl=64 time=0.109 ms
^C
--- 10.0.0.4 ping statistics ---
3 packets transmitted, 3 received, 0% packet loss, time 2029ms
rtt min/avg/max/mdev = 0.109/0.110/0.112/0.001 ms

To find out which interface to use:

(data-transfer) scarecrow@node1:~/data-transfer$ ifconfig

    eno1: flags=4099<UP,BROADCAST,MULTICAST>  mtu 1500
        ether 54:bf:64:6a:91:31  txqueuelen 1000  (Ethernet)
        RX packets 0  bytes 0 (0.0 B)
        RX errors 0  dropped 0  overruns 0  frame 0
        TX packets 0  bytes 0 (0.0 B)
        TX errors 0  dropped 0 overruns 0  carrier 0  collisions 0
        device interrupt 16  memory 0x92f00000-92f20000  

    # The one we used: Mellanox ConnectX-6 Dx (first port)
    -----------------
 -> ens4f0np0: flags=4163<UP,BROADCAST,RUNNING,MULTICAST>  mtu 1500
            inet 10.0.0.3  netmask 255.255.255.0  broadcast 10.0.0.255
            inet6 fe80::4e4d:abee:c38d:4b6f  prefixlen 64  scopeid 0x20<link>
            ether 10:70:fd:60:c1:cc  txqueuelen 1000  (Ethernet)
            RX packets 1021  bytes 63776 (63.7 KB)
            RX errors 0  dropped 0  overruns 0  frame 0
            TX packets 841  bytes 52495 (52.4 KB)
            TX errors 0  dropped 0 overruns 0  carrier 0  collisions 0
    -----------------

    ens4f1np1: flags=4163<UP,BROADCAST,RUNNING,MULTICAST>  mtu 1500
            inet 10.0.0.1  netmask 255.255.255.0  broadcast 10.0.0.255
            inet6 fe80::40ba:65b6:cb4e:4521  prefixlen 64  scopeid 0x20<link>
            ether 10:70:fd:60:c1:cd  txqueuelen 1000  (Ethernet)
            RX packets 36  bytes 4419 (4.4 KB)
            RX errors 0  dropped 0  overruns 0  frame 0
            TX packets 23  bytes 3072 (3.0 KB)
            TX errors 0  dropped 0 overruns 0  carrier 0  collisions 0

    lo: flags=73<UP,LOOPBACK,RUNNING>  mtu 65536
            inet 127.0.0.1  netmask 255.0.0.0
            inet6 ::1  prefixlen 128  scopeid 0x10<host>
            loop  txqueuelen 1000  (Local Loopback)
            RX packets 1782  bytes 168796 (168.7 KB)
            RX errors 0  dropped 0  overruns 0  frame 0
            TX packets 1782  bytes 168796 (168.7 KB)
            TX errors 0  dropped 0 overruns 0  carrier 0  collisions 0

    wlx04421a4d9e71: flags=4163<UP,BROADCAST,RUNNING,MULTICAST>  mtu 1500
            inet 81.236.101.159  netmask 255.255.254.0  broadcast 81.236.101.255
            inet6 fe80::7207:baa6:6cea:6bd3  prefixlen 64  scopeid 0x20<link>
            ether 04:42:1a:4d:9e:71  txqueuelen 1000  (Ethernet)
            RX packets 66840  bytes 48821412 (48.8 MB)
            RX errors 0  dropped 7  overruns 0  frame 0
            TX packets 61150  bytes 44974817 (44.9 MB)
            TX errors 0  dropped 0 overruns 0  carrier 0  collisions 0

If the connection is working building can be started

Building

Building the demo will create five executable files: run, web_gauge, fpga_emulator, transfer_real and transfer_synthetic.

  • run is the main data-transfer program.
  • fpga_emulator is a program to emulate incoming datastream (data generator) to send from one GPU to the other.
  • web_gauge is a web GUI showing a speedometer of the current transfer rate.
  • transfer_real is a demo utilizing incoming data from a socket, eg. fpga_emulator and displays the bandwidth to web_gauge.
  • transfer_synthetic showcase the maximum internal transferspeed within a program, ie. the program creates an array of ones, default is 2^26 bytes and writes to a remote region as fast as possible. The program web_gauge is also used to display the results.

To build all programs:

make

To only build fpga_emulator:

make fpga

To only build web_gauge:

make web_gauge

To cleanup and removal of executables:

make clean

Usage

Running the demo requires that both computers are connected via infiniband and correct software and modules are installed and running.

Receiver

Start the receiving end (Will receive a continuous flow of data)

./run --server -a <ADDRESS> -p <PORT>

Transmitter (Requires a running receiver)

Start the receiving end (Will receive a continuous flow of data)

./run --client -a <SERVER_ADDRESS> -p <PORT>

Data Generation (Requires a running transmitter)

Start the FPGA device sending data to a UDP socket or run the the emulator:

./fpga_emulator -a "localhost" -p <PORT>

Live Online Demo (Requires a running data-transfer)

Start the data-transfer demo:

./web_gauge --web-host <IP> -w <PORT> -a "localhost" -p 30000

View the running application running on specified IP and PORT:

firefox http://IP:PORT/

Problems and Known Issues

Network Cards Shutting Down

If the network adapters stops working after a period of time can be a result of insufficient cooling. The ConnectX cards require continuous cooling, we found that after exceeding 110°C, the modules were being unloaded from the system mlx5_core followed by many warnings when shown with dmesg. After exceeding the 120°C mark, the cards were physically shutdown by an onboard safety mechanism resulting in reloading the kernel modules was impossible and required a complete restart of the system.

From Nvidia https://docs.nvidia.com/networking/display/ConnectX5EN/Thermal+Sensors

The adapter card incorporates the ConnectX IC which operates in the range of temperatures between 0C and 105C.

There are three thermal threshold definitions which impact the overall system operation state:

    Warning – 105°C: On managed systems only: When the device crosses the 100°C threshold, a Warning Threshold message will be issued by the management SW, indicating to system administration that the card has crossed the Warning threshold. Note that this temperature threshold does not require nor lead to any action by hardware (such as adapter card shutdown).

    Critical – 115°C: When the device crosses this temperature, the firmware will automatically shut down the device.
    
    Emergency – 130°C: In case the firmware fails to shut down the device upon crossing the Critical threshold, the device will auto-shutdown upon crossing the Emergency (130°C) threshold.

The card's thermal sensors can be read through the system’s SMBus. The user can read these thermal sensors and adapt the system airflow in accordance with the readouts and the needs of the above-mentioned IC thermal requirements.

To check temperature of the cards install the

To find the cards:

(data-transfer) scarecrow@scarecrow:~/data-transfer$ lspci | grep Mellanox
04:00.0 Ethernet controller: Mellanox Technologies MT2892 Family [ConnectX-6 Dx]
04:00.1 Ethernet controller: Mellanox Technologies MT2892 Family [ConnectX-6 Dx]

To probe 04:00.0 (requires root privileges):

sudo mget_temp -d 04:00.0
53

Runtime Errors and Solutions

Aborted (core dumped) & UCX ERROR failed to register address

If during runtime the following error occurs a solution is to use smaller buffers. We are not entirely sure how this issue is emerging. We found that a buffer size greater than >115MB would result in the following errors:

(data-transfer) scarecrow@scarecrow:~/data-transfer$ ./transfer_synthetic --transmitter -p 12377 -sp 45004 -a 10.0.0.4  --n-bytes 1000000000
[1658921155.651848] [scarecrow:15571:0]          ib_log.c:254  UCX  ERROR ibv_reg_mr(address=0x7fe574000000, length=1000000000, access=0xf) failed: Bad address
[1658921155.651869] [scarecrow:15571:0]          ucp_mm.c:159  UCX  ERROR failed to register address 0x7fe574000000 mem_type bit 0x2 length 1000000000 on md[3]=mlx5_0: Input/output error (md reg_mem_types 0x3)
[1658921155.651874] [scarecrow:15571:0]     ucp_request.c:519  UCX  ERROR failed to register user buffer datatype 0x8 address 0x7fe574000000 len 1000000000: Input/output error
[scarecrow:15571:0:15571]        rndv.c:2378 Assertion `status == UCS_OK' failed
==== backtrace (tid:  15571) ====
0  /home/scarecrow/miniconda3/envs/data-transfer/lib/python3.7/site-packages/ucp/_libs/../../../../libucs.so.0(ucs_handle_error+0x2dc) [0x7fe71a27d7cc]
1  /home/scarecrow/miniconda3/envs/data-transfer/lib/python3.7/site-packages/ucp/_libs/../../../../libucs.so.0(ucs_fatal_error_message+0xb8) [0x7fe71a27a708]
2  /home/scarecrow/miniconda3/envs/data-transfer/lib/python3.7/site-packages/ucp/_libs/../../../../libucs.so.0(ucs_fatal_error_format+0x114) [0x7fe71a27a824]
3  /home/scarecrow/miniconda3/envs/data-transfer/lib/python3.7/site-packages/ucp/_libs/../../../../libucp.so.0(ucp_rndv_rtr_handler+0x608) [0x7fe71a331298]
4  /home/scarecrow/miniconda3/envs/data-transfer/lib/python3.7/site-packages/ucp/_libs/../../../../ucx/libuct_ib.so.0(+0x3eecd) [0x7fe6ef785ecd]
5  /home/scarecrow/miniconda3/envs/data-transfer/lib/python3.7/site-packages/ucp/_libs/../../../../libucp.so.0(ucp_worker_progress+0x6a) [0x7fe71a2fecca]
6  /home/scarecrow/miniconda3/envs/data-transfer/lib/python3.7/site-packages/ucp/_libs/ucx_api.cpython-37m-x86_64-linux-gnu.so(+0x290ba) [0x7fe71a3b70ba]
7  /home/scarecrow/miniconda3/envs/data-transfer/lib/libpython3.7m.so.1.0(_PyMethodDef_RawFastCallDict+0x9c) [0x7fe71ea68d7c]
8  /home/scarecrow/miniconda3/envs/data-transfer/lib/libpython3.7m.so.1.0(+0x1ed8ad) [0x7fe71ea698ad]
9  /home/scarecrow/miniconda3/envs/data-transfer/lib/libpython3.7m.so.1.0(PyObject_Call+0x64) [0x7fe71ea69c04]
10  /home/scarecrow/miniconda3/envs/data-transfer/lib/libpython3.7m.so.1.0(_PyEval_EvalFrameDefault+0x434a) [0x7fe71e8e4aea]
11  /home/scarecrow/miniconda3/envs/data-transfer/lib/libpython3.7m.so.1.0(_PyEval_EvalCodeWithName+0x920) [0x7fe71e99de80]
12  /home/scarecrow/miniconda3/envs/data-transfer/lib/libpython3.7m.so.1.0(_PyFunction_FastCallKeywords+0x8a) [0x7fe71ea690ea]
13  /home/scarecrow/miniconda3/envs/data-transfer/lib/libpython3.7m.so.1.0(+0x64416) [0x7fe71e8e0416]
14  /home/scarecrow/miniconda3/envs/data-transfer/lib/libpython3.7m.so.1.0(_PyEval_EvalFrameDefault+0x47a6) [0x7fe71e8e4f46]
15  /home/scarecrow/miniconda3/envs/data-transfer/lib/libpython3.7m.so.1.0(+0x6ddea) [0x7fe71e8e9dea]
16  /home/scarecrow/miniconda3/envs/data-transfer/lib/libpython3.7m.so.1.0(_PyFunction_FastCallDict+0x2e3) [0x7fe71ea694f3]
17  /home/scarecrow/miniconda3/envs/data-transfer/lib/libpython3.7m.so.1.0(_PyObject_Call_Prepend+0xf7) [0x7fe71ea6bc17]
18  /home/scarecrow/miniconda3/envs/data-transfer/lib/libpython3.7m.so.1.0(_PyObject_FastCallKeywords+0xd9) [0x7fe71ea69dd9]
19  /home/scarecrow/miniconda3/envs/data-transfer/lib/libpython3.7m.so.1.0(+0xf409b) [0x7fe71e97009b]
20  /home/scarecrow/miniconda3/envs/data-transfer/lib/libpython3.7m.so.1.0(_PyMethodDef_RawFastCallDict+0xf1) [0x7fe71ea68dd1]
21  /home/scarecrow/miniconda3/envs/data-transfer/lib/libpython3.7m.so.1.0(_PyCFunction_FastCallDict+0x29) [0x7fe71ea699d9]
22  /home/scarecrow/miniconda3/envs/data-transfer/lib/libpython3.7m.so.1.0(_PyEval_EvalFrameDefault+0x7e51) [0x7fe71e8e85f1]
23  /home/scarecrow/miniconda3/envs/data-transfer/lib/libpython3.7m.so.1.0(+0x6ddea) [0x7fe71e8e9dea]
24  /home/scarecrow/miniconda3/envs/data-transfer/lib/libpython3.7m.so.1.0(+0x64416) [0x7fe71e8e0416]
25  /home/scarecrow/miniconda3/envs/data-transfer/lib/libpython3.7m.so.1.0(_PyEval_EvalFrameDefault+0x47a6) [0x7fe71e8e4f46]
26  /home/scarecrow/miniconda3/envs/data-transfer/lib/libpython3.7m.so.1.0(+0x6ddea) [0x7fe71e8e9dea]
27  /home/scarecrow/miniconda3/envs/data-transfer/lib/libpython3.7m.so.1.0(+0x64416) [0x7fe71e8e0416]
28  /home/scarecrow/miniconda3/envs/data-transfer/lib/libpython3.7m.so.1.0(_PyEval_EvalFrameDefault+0x47a6) [0x7fe71e8e4f46]
29  /home/scarecrow/miniconda3/envs/data-transfer/lib/libpython3.7m.so.1.0(+0x6ddea) [0x7fe71e8e9dea]
30  /home/scarecrow/miniconda3/envs/data-transfer/lib/libpython3.7m.so.1.0(+0x64416) [0x7fe71e8e0416]
31  /home/scarecrow/miniconda3/envs/data-transfer/lib/libpython3.7m.so.1.0(_PyEval_EvalFrameDefault+0x47a6) [0x7fe71e8e4f46]
32  ./transfer_synthetic(+0xcb76) [0x55c388319b76]
33  ./transfer_synthetic(+0xed7e) [0x55c38831bd7e]
34  ./transfer_synthetic(+0x1050a) [0x55c38831d50a]
35  ./transfer_synthetic(+0x16e49) [0x55c388323e49]
36  ./transfer_synthetic(+0x10188) [0x55c38831d188]
37  ./transfer_synthetic(+0xae26) [0x55c388317e26]
38  /home/scarecrow/miniconda3/envs/data-transfer/lib/libpython3.7m.so.1.0(PyModule_ExecDef+0x4b) [0x7fe71ea2b3eb]
39  ./transfer_synthetic(+0xb709) [0x55c388318709]
40  ./transfer_synthetic(+0xbdaf) [0x55c388318daf]
41  /lib/x86_64-linux-gnu/libc.so.6(__libc_start_main+0xf3) [0x7fe71e6ae083]
42  ./transfer_synthetic(+0xbe91) [0x55c388318e91]
=================================
Aborted (core dumped)

ucp._libs.exceptions.UCXError: Device is busy

Current port is busy, terminate the process using it or use another port

UCX WARN transport 'ib' is not available

The networking library could not use Infiniband. A fix is to recompile UCX with Infiniband support, see building UCX for more information.

If running dmesg and the output looks like this:

(data-transfer) two-face@twoface:~/data-transfer$ dmesg
...
[18846.942527] port_module: 28 callbacks suppressed
[18846.942536] mlx5_core 0000:04:00.0: Port module event: module 0, Cable unplugged
[18846.944179] mlx5_core 0000:04:00.0: mlx5_pcie_event:293:(pid 4865): Detected insufficient power on the PCIe slot (26W).
[18846.944259] mlx5_core 0000:04:00.1: mlx5_pcie_event:293:(pid 4843): Detected insufficient power on the PCIe slot (26W).
[18846.989578] mlx5_core 0000:04:00.0 ens4f0np0: Link down
[18847.678446] mlx5_core 0000:04:00.1: Port module event: module 1, Cable unplugged
[18847.686809] mlx5_core 0000:04:00.1 ens4f1np1: Link down

The physical connection is down, make sure both computers are able to ping eachother before running the demo.

libpython3.7m.so.1.0 not found

If during runtime, the linker cannot find libpython3.7m.so.1.0 like this:

error while loading shared libraries: libpython3.7m.so.1.0: cannot open shared object file: No such file or directory

A temporal solution is to export the path to the module using:

export LD_LIBRARY_PATH=~/miniconda3/envs/data-transfer/lib/

About

Data-transfer between COTS devices using GPUDirect RDMA over UCX

cythonnvidiaucx-pyccudagpudirect-rdma

License:MIT License

Languages

Language:Cython 76.6%Language:Cuda 9.2%Language:C 8.3%Language:Makefile 3.0%Language:Python 2.2%Language:Shell 0.7%