PVM Howto
Contents
- Setting up pvm for python
- PVM commands
- Using remote windows with pvm
- Troubleshooting
- Enabling password-less login from A to B
Setting up pvm for python
Install pvm:
See the Install dependencies instructions or, if there are problems, refer to the troubleshooting howto Install pvm from source.
Note that you will need the header and library files for compiling the pypvm module. You get them automatically when installing from source but you need to select an additional pvm-devel package if you install from rpm.
Make sure you can log into all nodes without password:
See Enabling password-less login from A to B at the end of this document
Prepare environment:
export PVM_ROOT=/usr/lib/pvm3 export PVM_RSH=/usr/bin/ssh export PVM_ARCH=LINUX ## or LINUX64 or LINUXX86_64 on 64bit systems
pvm3 often rather resides in /usr/share/pvm3 -- adapt PVM_ROOT accordingly. Adjust $PVM_ARCH to the string returned by $PVM_ROOT/lib/pvmgetarch. So this should work more generally:
export PVM_ARCH=`$PVM_ROOT/lib/pvmgetarch` ## note the back-quotes!
Note: $PVM_ARCH is needed for compiling pypvm and should match a directory existing in $PVM_ROOT/lib. It may not be needed for running pvm except under some circumstances, e.g. if you use a different pvm binary folder (see next point). It is hence saver to always define it.
Adapt pvm binary folder:
PVM looks in the folder $PVM_ROOT/bin/$PVM_ARCH (for example /usr/share/pvm3/bin/LINUX) for executables. In our case, we need python and xterm. All the other programs used by biskit will be started from within python and PVM doesn't need to know about them.
If you don't have write-access to the system-wide $PVM_ROOT folder or want to use different binary versions, try creating a folder in your home (which is the default working directory of pvm):
cd ~ mkdir pvm3; mkdir pvm3/bin; mkdir pvm3/bin/$PVM_ARCH
Link the correct python version and xterm into this folder or into $PVM_ROOT/bin/$PVM_ARCH:
cd pvm3/bin/$PVM_ARCH ln -s `which python` . ln -s `which xterm` .
Note: the local binary folder in home seems not to work in practice. The best option is to adapt the $PVM_ROOT/bin/$PVM_ARCH folder directly. If really needed, make a complete copy of the pvm3 folder, point the $PVM_ROOT variable to this folder, adapt its bin/$PVM_ARCH folder and make sure that you indeed use $PVM_ROOT/lib/pvm for starting up the master (i.e. not the default one from which you made the copy). To avoid confusion, it is much better to have only one pvm folder installed on a node.
Repeat the above steps on every node to be used with pvm.
Test the raw pvm setup
start the pvm demon:
cd ~ ## important if you want to use a local ~/pvm3/bin folder pvm
Add a node (e.g. a computer called hal):
pvm> add hal add hal 1 successful HOST DTID hal 80000 pvm>
List all registered nodes (in this case, kaa is the local computer):
pvm> conf conf 2 hosts, 2 data formats HOST DTID ARCH SPEED DSIG kaa 40000 LINUX64 1000 0x00408c41 hal 80000 LINUX 1000 0x00408841
Spawn a test process ('kaa' is the name of the local host here):
pvm> spawn -kaa xterm spawn -kaa xterm 1 successful t40005 pvm>
You should see a xterminal popping up.
Spawn a test process on a remote host:
pvm> spawn -hal xterm spawn -hal xterm 1 successful t80001
If it looks like this, but you don't see the xterminal, you need to adapt the X-Permissions. See Using remote windows with pvm! If you get a "No such file" Error, you need to check your pvm binary path (e.g. pvm3/bin/LINUX64; make sure it exists and the link to xterm is correct).
Shut down pvm:
pvm> halt
In case things go wrong, make sure there are no .pvm* files left over in the /tmp folders of any node (at least none from yourself).
in Master.py: call pvmInit(hosts) with list of hosts
run Master.py (See example code in ExampleMaster.__main__)
stop pvm daemon (optional):
pvm pvm> halt
PVM commands
command | function |
add host_name | manually add a host |
remove host_name | remove it |
reset | kill all your running pvm processes |
halt | stop pvm altogether |
exit / ^D | exit the PVM shell without quitting the PVM demon |
Using remote windows with pvm
For security reasons, the default setting in recent linux distributions is to not allow a machine to remotely open a window. So, by default the X server on these machines is running as "tcp nolisten". The current implementation of getting remote windows from PVM slaves is using "xterm - display host:0.0" which will not run with the new secure settings.
To change back to the old insecure X server settings (example from a Linux box running the CentOS distriburion):
- Either comment in DisallowTCP=false in the file /etc/X11/gdm/gdm.conf
- Or use the graphical tools from menu System Settings -> Login Screen -> Security and uncheck: Always disallow TCP connections to the X server ...
Then on your local machine allow remote hosts to access the X server:
allow all hosts:
xhosts +
or only specific hosts:
xhost +remote.host.name
Finally test that it is working (you must have password free ssh set up):
ssh remote.host.name xterm -display local.host.name:0.0
An xterm window from the remote machine should show up on your local machine.
Troubleshooting
A slave can not be spawned:
- kill any pvm process left over on that node
- delete old pvm* files in /tmp on that node
- try again
Scripts that help you clean up:
- scripts/Biskit/killpp.py kills processes on many nodes
- scripts/Biskit/rm_pvm.py removes your pvm* from /tmp of all nodes
PVM doesn't use (for calculations) the node were the master is running:
- add this node with the full name e.g. otto.bis.pasteur.fr instead of otto
PVM modules cannot be imported:
Biskit gives a warning like this:
Warning (ignored): Could not import PVM modules. Please check that the pypvm_core library is compiled! Parallelisation is not available.
or like this:
exceptions.ImportError in |somename|.py line xx: ('No module named pypvm_core',).
- Make sure that you have installed pypvm (see Install dependencies).
- Check your $PYTHONPATH on each node and make sure your python is importing the Biskit and pypvm version it ought to.
PVM terminates after adding host
- Symptom:
pvm crashes after manually adding a host, for example:
pvm> add kaa add kaa [1] 11067 terminated pvm
From Biskit this may look like this:
PVM.ExampleMaster test ... adding 3 hosts to pvm... Process Python terminated
- Problem:
On Ubuntu, the pvm master node may identify itself with the localhost IP (127.0.0.1) which cannot be used by the slave nodes.
- Solution:
Edit /etc/hosts and put the actual (network) IP of the master node into the first line before 127.0.0.1. For example:
192.168.0.98 urfin 127.0.1.1 urfin 127.0.0.1 localhost
See also: http://stackoverflow.com/questions/2253354/pvm-terminates-after-adding-host
Enabling password-less login from A to B
- 3 steps:
- generate a identity file on A
- copy it over to B
- add this file to the list of authorized login
The following works for ssh2:
on A:
>>> ssh-keygen -t rsa
Hit return on each question (3)
>>> ssh-keygen -t dsa
Hit return on each question (3)
copy *pub files from A to B:
>>> cd ~/.ssh >>> scp id_dsa.pub "B:.ssh/id_dsa.pub-Bis" >>> scp id_rsa.pub "B:.ssh/id_rsa.pub-Bis"
on B "authorize" login from A:
>>> ssh B >>> cd $HOME/.ssh (you might need to create it ) >>> cat id_rsa.pub-Bis id_dsa.pub-Bis >> authorized_keys2
- Or download and use ssh-keyinstall to set up your system:
- http://www.stearns.org/ssh-keyinstall/
That's it. Test!
Still having problems? Check for firewalls!