Re: misc. openmosix questions.

From: "Kris Buytaert" <mlkb-1zPVpAQ/[email protected]>
To: "Michael C. Wu" <[email protected]>


> On Thu, 2006-04-06 at 11:44 -0500, Michael C. Wu wrote:
> Unless you are a kernelhacker or want to run on beta code you should go
> for 2.4 .. 2.6 is still in development.

Right, it says on the OM homepage, but I was wanting to verify that the 
webpage has been updated. :)

>> 2. Has anyone had OpenMOSIX automatically forward processes started by
>> Matlab  7?  Does it work?  Theoretically, since Matlab sees one big oM
>> machine, I only need licenses for users on the main node, right?
>>
> As documented extensiveley on the wiki people seem to have a mixed
> experience with getting mathlab to migrate,  the -nojvm option solves
> the failure to migrate however doesn't help on increasing performance as
> mathlab seems to be very system call intensive.

(Yeah, I read the wiki.)

From my personal trials, Matlab 7sp3 and 2006a do not work in standalone 
mode.
However, Matlab has this new Distributed Computing Toolbox and Distributed 
Computing Engine that allows you to basically do MPI, and offload jobs to 
various "worker nodes".  Each worker node, however, requires a worker 
license (Doesn't save you license fees).  The workers can migrate, but it 
doesn't save you any licenses.   Moreover, the workers still require JVM to 
be running, and I do not know if JVM is the reason that Matlab sends out 
system calls, thus requiring Matlab to run on the home node.

(How do I get this into the wiki?)

>> 3. Our computation requires constant disk I/O, what would be the 
>> recommended
>> way to connect to our fileserver in oM?  NFS? OMFS?
>>
>
> We removed oMFS from openMosix
> In 2.6 today I'd say some iSCSI setup, but as we are not there yet..

I setup channel bonding with NFS mounted on each node, but do not see if the 
slave oM nodes are running on local NFS or just getting data from network 
via oM internals.

>> 4. I plan on channel-bonding with 2 GigE for each node.  I only see a 
>> very
>> short page on this in the how-to.  Is there any trick to this, things to
>> watch out for?
>>
> Bonding depends on your switching infrastructure...  on the linux side
> it is easy.

(Will someone put this in the Wiki and the tutorial?)

The tutorial should be modified as such:

* Build bonding module, and do not build it into kernel statically.
modprobe bonding miimon=100
ifconfig bond0 up
ifenslave bond0 eth0 eth1
(Many distros come with their own ifenslave or do not have one, you should 
use the ifenslave that comes with your kernel source.  Read 
/usr/src/linux/Documentation/networking/bonding.txt )

ifconfig bond0 <ip> netmask <netmask> mtu <some number bigger than 1500>
===========
Gentoo has built-in channel bonding support in the RC scripts:
put the following two lines in /etc/modules.conf
alias bond0 bonding
options bond0 mode=0 miimon=100

put the following 2 lines in /etc/conf.d/net
slaves_bond0="eth0 eth1"
config_bond0=( "10.0.0.1 netmask 255.255.255.0 mtu 9000" )

as root:
emerge ifenslave
ln -s /etc/init.d/net.lo /etc/init.d/net.bond0
rc-update add net.bond0 default

After this, whenever you reboot gentoo, it will bring up the interfaces.

>> 5. The computers running as slave nodes will also be users' desktops. 
>> Can I
>> safely give them accounts locally to slave nodes?  If they start 
>> computation
>> on the slave node, the processes should migrate efficiently right?
> I don't really un derstand your question here..

The scenario is this:
1 oM master node
4 oM slave nodes

Users have local login (ssh) to slave nodes.  Can these users login to slave 
nodes and use it without knowing about oM processes?

I have since found out the answer is yes.

>> 6. I want to have checkpointing of processes.  It seems that there is
>> existing infrastructure.  Which package should I use?
>>
> chpox on openMosix 2.4

Slightly off-topic, I guess there is no real process checkpointing that 
works linux2.6 kernels.
Does anyone know of one?


Thanks Kris.

Regards to everyone,

Michael 




-------------------------------------------------------
Using Tomcat but need to do more? Need to support web services, security?
Get stuff done quickly with pre-integrated technology to make your job easier
Download IBM WebSphere Application Server v.1.0.1 based on Apache Geronimo
http://sel.as-us.falkag.net/sel?cmd=lnk&kid=120709&bid=263057&dat=121642