Spock Installation: Computing Node - calab-ntu/gpu-cluster GitHub Wiki

1. Check switche settings on MB

  1. VGA switch -> off image

  2. IPMI switch -> left (default) image

  3. PSU(PHANDEKS ) hybrid -> press down

  4. Change cooling fan header from CPU_OPT to CHA_FAN1 image

  5. Plug the micro USB plug off the CPU pump.

2. Set up BIOS

  1. Bootup machine with BIOS flash disk plugged in.

    If the machine is boot for the first time, it would ask if you want to initial the CPU config. Press Y to confirm initial.

  2. Check BIOS version an update

    • Get into BIOS with press delete or F2 during booting.
    1. Check BIOS version : Main -> BIOS Information -> Version

      if Version = 1106 x64, then skip the steps in Update BIOS

    2. Plug in the USB disk with "BIOS" label to USB socket labeled with "BIOS".
    3. Keep pressing delete or F2 during booting to get in to BIOS.
    4. Tool -> ASUS EZ Flash 3 Utility
    5. Find the folder PRO_WS_WRX80E-SAGE_SE_WIFI-ASUS-1106
    6. Find the file PRO-WS-WRX80E-SAGE-SE-WIFI-ASUS-1106.CAP
    7. Yes
    8. Reboot with save changes and exit or press F10.
    9. Check again.

  3. DRAM overclock setting

    1. Ai Tweaker -> Ai overclock Tuner -> Choose D.O.C.P
    2. D.O.C.P -> Choose D.O.C.P DDR4-3200 16-18-18-38-1.35V
    3. F10 reboot.
    4. Check: Main -> Total Memory : 262144 MB -> Speed : 3200 MHz

  4. Enable NUMA

    1. Advanced -> AMD CBS -> DF Common Option -> Memory Addressing
    2. NUMA nodes persocket -> Choose NPS2

  5. F10 Reboot

3. Install ubuntu server 22.04

  1. Download ubuntu 22.04 from https://www.ubuntu-tw.org/modules/tinyd0/ Make bootable USB disk with rufus.

  2. Set up boot disk in BIOS

    1. Boot with bootable USB disk plugged in.
    2. Get into BIOS with press delete or F2 during booting.
    3. Boot > Choose USB to boot.
    4. F10 to reboot.

  3. Install Ubuntu 22.04 0. Choose Try or install ubuntu server

    1. Select language : English -> Done

    2. Keyboard configuration.

      • Layout : English (US)
      • Variant : English (US) -> Done

    3. Choose type of install

      • Ubuntu Server
      • Search for third party drivers -> Done

    4. Network connecions -> Continue without network

    5. Configure proxy -> Done

      Leave the field empty.

    6. Configure Ubuntu archive mirror -> Done

      Don't change the url

    7. Guided storage configuration

      1. Custom storage layout -> Done
      2. Select disks and reformat all of them.
        1. /boot
          1. Choose free space -> Add GPT Partition
          • Size : 1G
          • Format: ext4
          • Mount: /boot -> Create
        2. / : Same steps as /boot with changes
          • Size : 448G
          • Mount : /
        3. swap
          • Size :

            leave empty to get rest of volume

          • Format : swap -> Done -> Continue

    8. Porfile setup

      • Your name: spock**

        ** is the number of node name

      • Your server's name: spock**
      • Pick a username: tmp_account
      • Choose a password: ************
      • Confirm your password: ***********

    9. Upgrade to Ubuntu Pro -> Skip Ubuntu Pro setup for now -> Done

    10. SSH Setup -> Done

      Don't check the option.

    11. Third-party drivers

      • Do not install third-party drivers now -> Done

    12. Reboot Now -> Unplug the install medium and press enter to reboot.

  4. Check.

    1. Kernel : uname -r -> 5.15.0-60-generic
    2. CPU : lscpu | grep Model -> AMD Ryzen Threadripper PRO 5975WX 32-Cores
    3. RAM : sudo dmidecode memory | grep Speed 
      Configured Speed: 3200 MT/s
Speed: 2667 MT/s
    4. NUMA : lscpu | grep NUMA 
      NUMA node(s)           : 2
NUMA node0 CPU(s) : 0-15, 32-47
NUMA node1 CPU(s) : 16-31, 48-63

4. Set up settings

  1. Network settings.

    1. Edit netplan : sudo vim /etc/netplan/00-installer-config.yaml 
      # This is the network config written by 'subiquity'
network:
  ethernets:
    enp*****0:
      dhcp4: true
    enp*****1:
      dhcp4: false
      addresses: [192.168.0.2**/22] # ** would be replaced by the number of node.
      nameservers:
        addresses: [140.112.254.4]
      routes:
        - to: default
          via: 192.168.0.1
version:2
    2. Apply netplan: sudo netplan apply
    3. Poweroff the machine and move it to machine room.
      • Plug the ethernet cable to the upper ethernet port.
    4. Check
      1. ip settings : ip addr show dev enp*****1 
        inet 192.168.0.2**/22
      2. DNS server : resolvectl status 
        Link 3 (enp*****1)
    Current Scopes: DNS
        Protocols: +DefaultRoute +LLMNR -mDNS -DNSOverTLS DNSSEC=no/unsupported
Current DNS Server: 140.112.254.4
       DNS Servers: 140.112.254.4
      3. ping 192.168.0.150
    5. Get system network informations. 
      sudo -i
scp [your_account]@192.168.0.150:/work1/shared/spock/etc/hosts /etc/hosts

  2. Update system 0. Operate in sudo privilage sudo -i

    1. apt update
    2. apt-get install -y linux-image-5.15.0-78-generic Press enter twice as kernel update UI appears.
    3. reboot
    4. sudo -i
    5. check : uname -r 5.15.0-78-generic # or above
    6. Change group name of ID 1000 : groupmod --new-name calab tmp_account
    7. Set root password : passwd
    8. Delete /home/tmp_account : rm -r /home/tmp_account
    9. Change sh link from dash to bash: 
      sudo dpkg-reconfigure dash
# Then configure UI will ask if want to set /usr/bin/sh to dash
# Press "No" to set the /usr/bin/sh to bash
    10. apt install --only-upgrade -y openssh* wget

  3. Time stamp of command history

    1. su
    2. Add export HISTTIMEFORMAT='%d/%m/%y %T ' to the end of file /etc/profile
    3. source /etc/profile
    4. Check by history

  4. Set timezone

    1. su
    2. timedatectl set-timezone Asia/Taipei
    3. Check timedatectl show

  5. NFS settings

    1. Client
      1. sudo -i
      2. Install NFS client. apt -y install nfs-common
      3. Get auto mount settings from work1. 
        ssh [your_account]@eureka00 cat /work1/shared/spock/etc/fstab >> /etc/fstab
      4. Create directories. mkdir /software /work1 /projectV /projectW /projectX /projectY /projectZ
      5. Check the accessibility of the target NFS servers 
        showmount -e spock00     # /software 192.168.0.0/24 **[Skip on login node]**
showmount -e tumaz        # /home 192.168.0.0/24
showmount -e ironman    # /volume1/gpucluster1 192.168.0.0/24
                                          # /volume3/gpucluster3 192.168.0.0/24
showmount -e eater         # /volume1/gpucluster3 192.168.0.0/24
                                          # /volume2/gpucluster4 192.168.0.0/24
                                          # /volume3/gpucluster6 192.168.0.0/24
showmount -e pacific       # /volume1/gpucluster1 192.168.0.0/24
      6. Mount all remote directories. 
        mount /software; # Skip in process on login node 
mount /home; mount /work1; mount /projectW; mount /projectX; mount /projectY; mount /projectZ; mount /projectV
      7. Check : df -h 
        tumaz:/home                   208G   22G  176G  12% /home

ironman:/volume1/gpucluster1   70T   47T   24T  67% /work1
ironman:/volume3/gpucluster3   70T   70T  643G 100% /projectX

eater:/volume1/gpucluster3     70T   67T  3.6T  95% /projectY
eater:/volume2/gpucluster4     88T   77T   12T  88% /projectZ
eater:/volume3/gpucluster6     88T   75T   13T  86% /projectW

pacific:/volume1/gpucluster1  140T   20T  120T  15% /projectV

  6. NIS settings

    1. Install NIS client. sudo apt -y install nis
    2. Configure as a NIS Client.
      1. vim /etc/yp.conf , add follow text at the end. 
        domain tumaz.gpucluster.calab server tumaz
      2. vim /etc/nsswitch.conf 
        passwd:         files systemd nis
group:          files systemd nis
shadow:         files nis
hosts:          files dns nis
      3. Set NIS domain name, vim /etc/defaultdomain 
        tumaz.gpucluster.calab
      4. Start and enable nis. 
        systemctl restart ypbind
systemctl enable ypbind
    3. check :
      1. ll /home
      2. yptest : 1 test fail
      3. ypwhich : tumaz
    4. Logout and login with your own account su Delete tmp_account : userdel --remove tmp_account It's okay to receive error message: 
      userdel: tmp_account mail spool (/var/mail/tmp_account) not found 
userdel: tmp_account home directory (/home/tmp_account) not found

  7. Install GPU driver

    1. Set the text mode as default (since the NVIDIA driver cannot be installed while X window is running) systemctl set-default multi-user.target
    2. Reboot.
    3. su
    4. Install dkms : apt -y install dkms
    5. Disable nouveau : Create file /etc/modprobe.d/blacklist-nouveau.conf with content: 
      blacklist nouveau
options nouveau modeset=0
    6. Apply system changes update-initramfs -u
    7. Reboot.
    8. su
    9. Check nouveau is disabled : lsmod | grep nouveau

      This should print nothing.

    10. Install nvidia dirver
      1. Install : 
        su
sh /work1/shared/spock/package/cuda/cuda_12.1.0_530.30.02_linux.run --silent --driver
      2. Validate with cat /proc/driver/nvidia/version: 
        NVRM version: NVIDIA UNIX x86_64 Kernel Module  530.30.02
GCC version:  gcc version 11.3.0 (Ubuntu 11.3.0-1ubuntu1~22.04)
      3. Copy the default profile files. 
        cp /work1/shared/spock/init_script/*.sh  /etc/profile.d/
cp /work1/shared/spock/init_script/*.csh /etc/profile.d/

cp /work1/shared/spock/etc/rc.local /etc/
chmod +x /etc/rc.local
      4. Reboot
    11. Check nvidia-smi 
      NVIDIA-SMI 530.30.02              Driver Version: 530.30.02    CUDA Version: 12.1

  8. NTP client 0. su

    1. apt -y install ntp ntpdate
    2. Edit /etc/ntp.conf
      • Add pool time.google.com iburst
      • Comment out other pool servers.
    3. systemctl restart ntp
    4. systemctl status ntp
    5. systemctl enable ntp

  9. TORQUE

    1. Install the required packages 
      apt -y install libnuma-dev
apt -y install tcl-dev tk-dev
apt -y install libntirpc-dev
sh /work1/shared/spock/package/torque/src/torque-3.0.6/spock_library_set.sh
    2. Compile and install from source code. 
      cd /work1/shared/spock/package/torque/src/torque-3.0.6
# WARNING: do NOT run "spock_Install.sh" in parallel (i.e., install one node at a time)
# [Login node ] uncomment "--enable-server"
# [Computing nodes] comment "--enable-server"
sh spock_Install.sh >& log.spockXX
cd ../../etc
cp pbs_spock /etc/init.d/pbs
ln -s /etc/init.d/pbs /etc/systemd/system/

cp pbs.conf /etc/
# [Login node only]: edit "pbs.conf" to set "start_server=1" and "start_mom=0"
cp nodes_spock /var/spool/TORQUE/server_priv/nodes
systemctl enable pbs 

source /etc/profile.d/torque.sh
cd ../src/torque-3.0.6/
./torque.setup root
killall pbs_server
systemctl start pbs # This error message is fine: "LOG_ERROR::No such file or directory (2) in read_config, fstat: config"
systemctl status pbs
    3. Check cat /var/spool/TORQUE/pbs_environment : LANG=en_US.utf-8
    4. Setup overcommit-ratio and Disable overcommit-memory in crontab
      1. cp /work1/shared/spock/helper_script/disable_memory_overcommit.sh /root/
      2. Edit crontab with crontab -e and add a new line: 
        @reboot /usr/bin/sh /root/disable_memory_overcommit.sh 1> /tmp/disable_memory_overcommit.log 2>&1
    5. Setup prologue and epilogue to /var/spool/TORUE/mom_priv/ 
      cp /work1/shared/spock/package/torque/mom_priv/* /var/spool/TORQUE/mom_priv/

  10. InfiniBand

    ref. https://docs.nvidia.com/networking/display/MLNXOSv3105002/Getting+Started#heading-RerunningtheWizard

    1. Check hardware lspci | grep Mellanox 
      01:00.0 Infiniband controller: Mellanox Technologies MT28908 Family [ConnectX-6]
    2. Install necessary package 
      apt -y install libsasl2-dev  libldap2-dev libssl-dev
    3. Install driver
      1. su
      2. cd /work1/shared/spock/package/ib/adaptor/driver/MLNX_OFED_LINUX-5.9-0.5.6.0-ubuntu22.04-x86_64
      3. ./mlnxofedinstall 
        Device #1:
----------
Device Type:      ConnectX6
Part Number:      MCX653105A-HDA_Ax
Description:      ConnectX-6 VPI adapter card; HDR IB (200Gb/s) and 200GbE; single-port QSFP56; PCIe4.0 x16; tall bracket; ROHS R6
PSID:             MT_0000000223
PCI Device Name:  01:00.0
Base GUID:        0c42a10300ef2a1a
Versions:         Current        Available
FW             20.34.1002     20.36.1010
PXE            3.6.0700       3.6.0901
UEFI           14.27.0014     14.29.0014
Status:           Up to date
---------
      4. /etc/init.d/openibd restart
      5. reboot
    4. Check 0. su
      1. ibstatus 
        Infiniband device 'mlx5_0' port 1 status:
        default gid:     fe80:0000:0000:0000:0c42:a103:00ef:2a1a
        base lid:        0xffff
        sm lid:          0x0
        state:           4: ACTIVE
        phys state:      5: LinkUp
        rate:            200 Gb/sec (4X HDR)
        link_layer:      InfiniBand
      2. cat /etc/security/limits.conf 
        * soft memlock unlimited
* hard memlock unlimited
      3. systemctl status openibd 
        Active: active (exited)
      4. systemctl is-enabled openibd 
        enabled
      5. systemctl status opensmd 
        Active: inactive (dead)
      6. systemctl is-enabled opensmd 
        disabled
      7. hca_self_test.ofed 
        ---- Performing Adapter Device Self Test ----
Number of CAs Detected ................. 1
PCI Device Check ....................... PASS
Kernel Arch ............................ x86_64
Host Driver Version .................... MLNX_OFED_LINUX-5.9-0.5.6.0 (OFED-5.9-0.5.6): 5.15.0-69-generic
Host Driver RPM Check .................. PASS
Firmware on CA #0 HCA .................. v20.36.1010
Host Driver Initialization ............. PASS
Number of CA Ports Active .............. 1
Port State of Port #1 on CA #0 (HCA).....  UP 4X HDR (InfiniBand)
Error Counter Check on CA #0 (HCA)...... PASS
Kernel Syslog Check .................... PASS
Node GUID on CA #0 (HCA) ............... 0c:42:a1:03:00:ef:2a:1a
------------------ DONE ---------------------
      8. ibdev2netdev -v | grep -i MCX 
        0000:01:00.0 mlx5_0 (MT4123 - MCX653105A-HDAT) ConnectX-6 VPI adapter card, HDR IB (200Gb/s) and 200GbE, single-port QSFP56
                                                                                                           fw 20.36.1010 port 1 (ACTIVE) ==> ibp1s0 (Down)
      9. IB connection and band width test.
        1. Computing nodes -> Login node On spock00 
          ib_write_bw -aF
          On spockXX 
          ib_write_bw -aF spock00

          ************************************
* Waiting for client to connect... *
************************************
---------------------------------------------------------------------------------------
RDMA_Write BW Test
Dual-port       : OFF          Device         : mlx5_0
Number of qps   : 1            Transport type : IB
Connection type : RC           Using SRQ      : OFF
PCIe relax order: ON
ibv_wr* API     : ON
CQ Moderation   : 100
Mtu             : 4096[B]
Link type       : IB
Max inline data : 0[B]
rdma_cm QPs     : OFF
Data ex. method : Ethernet
---------------------------------------------------------------------------------------
local address: LID 0x02 QPN 0x0027 PSN 0xcb8c4 RKey 0x1fffbe VAddr 0x007f9c96aaa000
remote address: LID 0x01 QPN 0x0027 PSN 0x560b74 RKey 0x1fffbe VAddr 0x007f0894517000
---------------------------------------------------------------------------------------
#bytes     #iterations    BW peak[MB/sec]    BW average[MB/sec]   MsgRate[Mpps]
8388608    5000             23452.55            23452.55                  0.002932
---------------------------------------------------------------------------------------
        2. Computing nodes <- Login node On spock00 
          ib_read_bw -aF
          On spockXX 
          ib_read_bw -aF spock00

          ************************************
* Waiting for client to connect... *
************************************
---------------------------------------------------------------------------------------
RDMA_Read BW Test
Dual-port       : OFF          Device         : mlx5_0
Number of qps   : 1            Transport type : IB
Connection type : RC           Using SRQ      : OFF
PCIe relax order: ON
ibv_wr* API     : ON
CQ Moderation   : 100
Mtu             : 4096[B]
Link type       : IB
Outstand reads  : 16
rdma_cm QPs     : OFF
Data ex. method : Ethernet
---------------------------------------------------------------------------------------
local address: LID 0x02 QPN 0x0028 PSN 0x593c01 OUT 0x10 RKey 0x1fffbf VAddr 0x007efc3f67f000
remote address: LID 0x01 QPN 0x0028 PSN 0xbaa0aa OUT 0x10 RKey 0x1fffbf VAddr 0x007f6fd2a85000
---------------------------------------------------------------------------------------
#bytes     #iterations    BW peak[MB/sec]    BW average[MB/sec]   MsgRate[Mpps]
8388608    1000             23517.75            23517.73                  0.002940
---------------------------------------------------------------------------------------
    5. Start mst to make us enable monitor IB adaptor 
      systemctl enable mst
systemctl start mst
mst status

  11. ssh without password for the root

    echo "#Allow only following users access this node by ssh" >> /etc/ssh/sshd_config
echo "AllowUsers root" >> /etc/ssh/sshd_config
cd /work1/shared/spock/ssh_root/
cp authorized_keys id_rsa* /root/.ssh/

# Verification
ssh spock00   # "yes" to "continue connecting"
ssh spockXX   # "yes" to "continue connecting"
exit
exit

5. install compilers [Login node only]

  1. Intel compiler
    1. su
    2. cd /opt
    3. ln -s /software/intel

6. install packages

  1. Module

    cd /work1/shared/spock/package/module/modules-5.1.1
make clean
./configure
make
make install

    After installation

    cp init/profile.sh /etc/profile.d/10-modules.sh
cp init/profile.csh /etc/profile.d/modules.csh
source init/bash

    Add /software/intel/oneapi/modulefiles to default module directories by adding the line to the file /usr/local/Modules/etc/initrc

    module use /software/intel/oneapi/modulefiles

    Set up preload module

    ln -s /software/modulefiles/default_modules.sh /etc/profile.d/default_modules.sh

  2. python2

    source /etc/profile.d/openmpi.sh; source /etc/profile.d/intel.sh; source /etc/profile.d/hdf5.sh
apt -y install python2 python2-dev
apt -y install python-tk
cd /work1/shared/spock/package/python2
python2 get-pip.py
sh install-python-packages.sh

  3. python3

    apt -y install python3 python3-dev
apt -y install python3-tk
apt -y install python3-pip
cd /work1/shared/spock/package/python3
sh install-python-packages.sh

    Add /usr/local/bin to PATH by adding a line at the end of /etc/profile

    export PATH=/usr/local/bin:$PATH

7. Miscellaneous setup

  1. IPMI tool

    1. Install IPMI driver and tool : apt -y install openipmi ipmitool
    2. Check : ipmitool sensor get "CPU Temp."

  2. ffmpeg apt -y install ffmpeg

  3. gnuplot apt -y install gnuplot-x11

  4. screen apt -y install screen

  5. pdsh apt -y install pdsh

  6. locate apt -y install plocate

  7. ClamAV

    apt -y install clamav clamav-daemon
systemctl stop clamav-freshclam
freshclam
systemctl start clamav-freshclam
systemctl enable clamav-freshclam

  8. X11 server

    apt -y install xorg openbox

  9. CPU usage monitor

    apt -y install sysstat

  10. Image display feh

    apt -y install feh

  11. cmake

    apt -y install cmake

  12. GNU Readline Library

    apt -y install lib32readline8 lib32readline-dev

  13. Disable auto update.

    1. Edit the apt config file at /etc/apt/apt.conf.d/20auto-upgrades as follow. 
      APT::Periodic::Update-Package-Lists "0";
APT::Periodic::Unattended-Upgrade "0";
    2. Apply config 
      apt-config dump APT::Periodic::Update-Package-Lists
apt-config dump APT::Periodic::Unattended-Upgrade

8. Check

  1. CPU burn-in test

    1. Install CPU test program 
      apt -y install stress-ng
    2. Run CPU test 
      stress-ng --cpu 0 --timeout 30m &
    3. Detect CPU temperature every minute during test 
      for i in {1..40}; do ipmitool sensor | grep "CPU Temp."; sleep 1m; done
      AMD Threadripper allows temperature up to 95 degree. And the non-critical upper limit is 85 degree. for spock02 the highest temperature is 82 degree.

  2. GPU burn-in test

    cd /work1/shared/spock/tests/gpu_burn-in/gpu-burn
./gpu_burn 1800 # run for 30 minutes

    during the test, watch the gpu temperature shown on screen. For RTX3080Ti, hightest temperature is 93 degree celsius. And the non-critical upper limit is 90 degree. For spock02, the highest temperature is 81 degree.

  3. MPI suit test [Run as regular user]

    1. Download @spock00 git clone https://github.com/open-mpi/mpi-test-suite.git
    2. Compile @spock00 cd mpi-test-suite ./autogen.sh ./configure CC=mpicc make
    3. Run tests cp /work1/shared/tests/mpi_test_suite/run_test.sh ./ qsub -I -lnodes=spockXX:ppn=32 cd {directory of mpi_test_suite} sh run_test.sh >& spockXX.log
    4. Check test result tail spockXX.log # Number of failed tests: 0