Cortex/Archive

From ISRWiki
Revision as of 14:01, 14 July 2017 by Giovanni Saponaro (talk | contribs) (→‎Customization: auto mounting of blackhole)
(diff) ← Older revision | Latest revision (diff) | Newer revision → (diff)
Jump to navigation Jump to search

Note: these methods are obsolete and kept here for historic reference only. Most probably, you may ignore this page and go back to Cortex.

Specifications

During 2009-2012, the setup was the following.

From an end-user perspective:

  • each machine hosts an Intel Pentium Dual-Core processor (2 x E2180 @ 2.00GHz);
  • memory: 2GB for each machine.

More specifically:

  • each of the client machines has the following specs: Intel Desktop DG965SS motherboard (info1, info2); Intel DualCore E2180 CPU @ 2 GHz with cache 1 MB; 2 x RAM 1GB DDR2 677 MHz; rackmount 19" 2U case from Chieftec (UNC-210S-B), with ATX 300W power supply;
  • the server machine has similar specs plus Intel PRO 1000GTLP/ENet gigabit ethernet card; 2 x WD2500JS 250GB SATAII 8MB disks for software RAID-1.

The Cortex computation rack actually contains 7 machines:

  • 1 server that manages startup, shutdown and the file system of the clients;
  • 6 clients (named cortex1...cortex6) that run user processes.

All clients numbered 1 to 5 mount the same file system. Therefore, performing changes in the file system of cortex[1-5] will reflect to all other four clients. Beware, though, that because of the way the file systems are mounted, there is some caching going on. This improves disk access performance, but strange phenomena might happen, i.e., after a file is modified and saved on one client, other clients can continue to see the old version of it for some time (probably less than one minute). The client cortex6 is separate for now, because it runs a 64 bit Linux.

For further details, see the Detailed setup section below.

Detailed setup

Network setup

Connectivity

Cortex machines are connected to Cortex Switch, that links to VisLab switch with a fiber optic connection of 4Gbit/s.

Cortex nodes

Cortex server and clients have the following IPs and domain names:

  • Server: 10.10.1.240, server.visnet
  • Client 1: 10.10.1.1, cortex1.visnet
  • Client 2: 10.10.1.2, cortex2.visnet
  • Client 3: 10.10.1.3, cortex3.visnet
  • Client 4: 10.10.1.4, cortex4.visnet
  • Client 5: 10.10.1.5, cortex5.visnet
  • Client 6: 10.10.1.6, cortex6.visnet

For further details, see VisLab network.

Boot procedure

The clients boot via the network, using the PXE system. Each machine determines its own identity and asks the server for a kernel image and an initial ram disk. Kernel images and initial ram disks are stored on the server in the /tftpboot/ directory. Kernels can be stock kernels, but the initial ram disk must be created in a way that enables booting from the network. This is not as bad as it sounds. It involves invoking the command mkinitramfs.

The server decides which kernel and initramdisk to send to each machine based on the information stored in the two files: /tftpboot/pxelinux.cfg/default(for cortexes 1-5) and /tftpboot/pxelinux.cfg/01-00-19-d1-9e-e9-53 (for cortex6).

The roots and the homes of the users are also stored on the server machine, so they are mounted by each client at boot time.

Mounting of root directory

We are not sure which mechanism mounts the root filesystem, exactly. Here is the relevant line from /etc/fstab: 

 # <file system> <mount point>   <type>  <options>       <dump>  <pass>
 /dev/nfs        /               nfs     defaults        1       1

Mounting of home directory

The home directory is mounted via the Upstart system a few seconds after booting. When rebooting the system, it is possible to login while /home is still not mounted. In that case, log out and log in again, in order for your environment variables to be set correctly.

In November 2010, we created a file called /etc/init/mountHome-net.conf containing: 

 description     "Mount network filesystems"
 start on started networking or runlevel 2
 exec /usr/local/bin/mountHome.sh

and /usr/local/bin/mountHome.sh containing: 

 MOUNTED=$(mount | grep home)
 while [ -z "$MOUNTED" ]
 do
       su icub -c 'mount /home' &> /var/tmp/mountHomeUpstartOut.txt
       MOUNTED=$(mount | grep home)
       echo $MOUNTED
       sleep 1
 done
 ls /home > /var/tmp/mountHomeUpstartLs.txt
 runlevel > /var/tmp/mountHomeUpstartRunlevel.txt

After a successful mount, we should see something like 

 $ mount | grep home
 10.10.1.240:/nfsroot.home on /home type nfs (rw,user=icub,addr=10.10.1.240)

If /home is wrongly mounted with the noexec flag, users won't be able to execute binaries located inside it.

Server machine

The server has:

  • a boot folder for the clients at /tftpboot/pxelinux.cfg. It contains the files:
    • default - default boot file;
    • <mac_address> - specific for a machine with the given mac address.
  • startup scripts for each machine at /nfsroot/app

Client machines

The clients have:

  • A superuser account (compurack) to administer system-wide settings (configurations, libraries, etc.)
  • Normal user accounts. By default, the login script runs the contents of file $HOME/.bash_env, where users can set their environment variables, e.g., export ICUB_ROOT=$HOME/iCub. This works for both interactive shell sessions and non-interactive ones (i.e., commands remotely invoked by yarprun).
  • A yarp account to update and install the YARP library. Variable YARP_ROOT is set by default to /home/yarp/yarp2 for all users (in /etc/bash.bashrc) <-- change this policy
  • An icub account with sudo privileges (created with sudo adduser icub admin on 2009-06-30) <-- change this policy
  • cortex6's /etc/hosts file can include the following line:
 127.0.0.1	cortex6

Additional setup

Network tuning

  sysctl -w net.core.rmem_max=8388608
  sysctl -w net.core.wmem_max=8388608
  sysctl -w net.core.rmem_default=65536
  sysctl -w net.core.wmem_default=65536
  sysctl -w net.ipv4.tcp_rmem='4096 87380 8388608'
  sysctl -w net.ipv4.tcp_wmem='4096 65536 8388608'
  sysctl -w net.ipv4.tcp_mem='8388608 8388608 8388608'
  sysctl -w net.ipv4.route.flush=1

Other libraries, installed with Ubuntu packages

These packages were installed with sudo apt-get install: 

 libgtk2.0-dev libgtkmm-2.4-dev libglademm-2.4-dev glew-utils libglew1.5-dev libglut-dev git-core

OpenCV: 

 libcv-dev libhighgui-dev libcvaux-dev

System-wide libraries and repositories

YARP

Presently (November 2010), the yarp2 SVN repository is installed under user yarp (with sudo make install), last updated on 13-July-2010. <-- change this policy?

iCub

Presently (November 2010), the iCub SVN repository is installed under user icub (with sudo make install), last updated on 13-July-2010. <-- change this policy?

There was a conflict with iKin, which could not find libipopt.so.0, but it is now fixed thanks to setting the environment variable 

 LD_LIBRARY_PATH=/opt/Ipopt-3.5.5-linux-x86_32-gcc4.2.4/lib/

into /home/icub/.bash_env.

One module has been disabled in the CMakeList.txt file, because it was not compiling properly: crawling.

User repositories

you should add <iCub>/bin to your PATH by editing your ~/.bashrc like this: 

 PATH=$PATH:~/iCub/bin/
 ICUB_DIR=~/iCub/ <-- needs to change
 export ICUB_DIR
 ICUB_ROOT=$ICUB_DIR
 export ICUB_ROOT

You should also edit ~/.bash_env adding these lines: 

 export ICUB_DIR=$HOME/iCub <-- needs to change
 export ICUB_ROOT=$ICUB_DIR

this is needed when you connect non-interactively via ssh to a Cortex computer, for instance when execute a "yarp run ..." on a Cortex, from Chico2.

Be aware that Ubuntu 7.10 (the version currently installed on the cluster) has a conflict with iKin, specifically with iCub/conf/FindIPOPT.cmake (used by iKin): for now, in order to compile iKin, change the following line of FindIPOPT.cmake from 

  SET(IPOPT_LIB   ${IPOPT_LIB} gfortranbegin gfortran)

to 

  SET(IPOPT_LIB   ${IPOPT_LIB} gfortran)

Customization

Prompt ($PS1)

The prompt is set to user@cortex?:pwd$ in /etc/bash.bashrc. With those settings, if you log in to Cortex1, the prompt will be user@cortex1:~$. We chose to do so because sometimes it's convenient to have the number of the Cortex machine you're working on embedded in the prompt. By default, though, this configuration is overridden in the users' ~/.bashrc file, and the prompt is set to user@source regardless of the Cortex machine you log in to.
If you want to inhibit this behaviour in ~/.bashrc and thus have a prompt like user@cortex?:pwd, just comment these lines in your ~/.bashrc: 

  # set a fancy prompt (non-color, unless we know we "want" color)
  case "$TERM" in
  xterm-color)
      PS1='${debian_chroot:+($debian_chroot)}\[\033[01;32m\]\u@\h\[\033[00m\]:\[\033[01;34m\]\w\[\033[00m\]\$ '
      ;;
  *)
      PS1='${debian_chroot:+($debian_chroot)}\u@\h:\w\$ '
      ;;
  esac

However, for users created after 2009-05-07, the prompt is already set to user@cortex?:pwd$ by default.

Mounting network folders

The folder "database" on BlackHole is mounted automatically to /media/blackhole. For details, here is the content of /etc/fstab: 

 #MATTEO: removed because it wouldn't work at startup: file systems are mounted before the network connection is available.
 #//10.0.3.118/database /media/blackhole smbfs username=vislab,password=vislab,iocharset=utf8,sec=ntlm  0  0

Helper commands

These generic Linux commands should be written somewhere else, as they are not Cortex-specific

  • Check the kernel: uname -m
  • Check the file versions: file
  • Set bash shell in /etc/passwd
  • Check disk space: du –sh /home
  • Check per-user processes: ps -U <user>
Retrieved from "http://mediawiki.isr.tecnico.ulisboa.pt/index.php?title=Cortex/Archive&oldid=6787"