[GE users] Multiple SGEs

Mikhail Kuzminsky kus at free.net
Wed Jan 12 20:38:46 GMT 2011


According to Andreas Haas
> 
> Hi Ralph,
> 
> I agree use of multiple independent cells is a good mode to start with.
> Managing multiple independent cells is supported by Grid Engine ...
  We also plan to start the work w/multiple SGE's/multiple cells.
But I want to extend the question: will be "interoperable" the
configuration w/different SGE versions, for example SGE 5.3 & SGE 6.0 ?

Mikhail Kuzminsky
Zelinsky Institute of Organic Chemistry
Moscow

> well except with the end user interfaces (qsub/...) we failed to go
> the last mile so far.
> 
> My thinking is that having the ability to clearly specify a particular
> cell at the command line rather than trough env vars surely would help
> end users adoption. Please find idea's how to adress this by providing
> corresponding wrappers adding a -cell switch to qsub/.. interface
> 
>    http://gridengine.sunsource.net/servlets/ReadMsg?msgId=3972&listName=users
> 
> would be interesting to hear how it goes if you chose that approach.
> 
> Cheers,
> Andreas
> 
> On Thu, 25 Mar 2004, Ralph Castain wrote:
> 
> > Hi Ron
> >
> > Thanks for the info on the SGE_CELL environmental variable - we'll check
> > that out, and try to see if we can qsub to different SGE masters from a
> > single host.
> >
> > Your question about cells vs mult-clustering was intriguing. I think
> > that perhaps this is a semantic issue, so let me clarify our thinking.
> > We distinguish those two terms as:
> >
> > Grid cell: a group of machines, connected into a "grid" via high latency
> > network connections (e.g., TCP/IP over standard Ethernet), coordinated
> > by a "master" that handles the queues for each machine
> >
> > Cluster: a group of machines, connected via low latency interconnects
> > (e.g., Infiniband), coordinated by a "master" that handles the queues
> > for each machine
> >
> > In other words, the only difference between a multi-grid cell
> > installation and a multi-cluster installation is the latency of the
> > individual groups of machines. In our situation, we consider the system
> > to be of a mult-grid cell nature. All of our "groups" are of the grid
> > variety. Unfortunately, we cannot simply link all of the machines
> > together into one large "grid" because each grid cell is independently
> > owned and operated, with its own security system, job scheduling
> > priorities, etc.
> >
> > The only thing all the grid cells have in common (at the moment) is the
> > use of SGE as their "master". What we have managed to do is negotiate
> > with each grid cell owner the right to allow people from outside their
> > cell (but within the university!) to have a very tightly limited ability
> > to execute applications on their grid cell. Hence my questions regarding
> > the use of qsub to multiple SGE's.
> >
> > I'm sure someone out there will suggest using Globus as the "glue" to
> > tie this together. With all due respect to the Globus advocates, please
> > spare the electrons - we do not wish to expend the effort required on
> > Globus at this stage. First, we'd just like to get everyone used to the
> > idea of "sharing" access to their local grid resources, and validate
> > that there is in fact a benefit to more global access to resources.
> >
> > Hope that helps explain the situation. Feel free to provide any
> > suggestions on easier ways to accomplish our objective!
> >
> > Thanks
> > Ralph
> >
> > On Wed, 2004-03-24 at 22:30, Ron Chen wrote:
> >
> > > --- Ralph Castain <ralph.castain at colostate.edu> wrote:
> > > > Questions:
> > > >
> > > > (a) how does qsub know which GE to send the request
> > > > to?
> > >
> > > It looks at the $SGE_CELL variable to determine which
> > > sge cell you are currently pointing to. (the default
> > > cell is "default")
> > >
> > > > (b) If I do that multiple times, once for each GE
> > > grid
> > > > cell, will they overwrite each other?
> > >
> > > SGE cells do not communicate with each other, they are
> > > like seperate clusters.
> > >
> > > If you configure host X to be the submit host for cell
> > > A, since cells are seperate clusters, cell B will not
> > > know about host X, so you will need to do it again.
> > >
> > > I am not 100% sure, but you can try to configure a
> > > host to be submit host for a cell, and then point to
> > > another cell and try to submit a job.
> > >
> > > I think most people here don't use cells, there are
> > > actually several ways to do multi-clustering -- but of
> > > course we need to know why you use cells in the first
> > > place!
> > >
> > >  -Ron
> > >
> > >
> > > I'm not sure if/what/where the
> > > > script writes
> > > > anything, other than the environment-setting script
> > > > "settings.sh".
> > > >
> > > > Thanks
> > > > Ralph
> > > >
> > > >
> > >
> > >
> > > __________________________________
> > > Do you Yahoo!?
> > > Yahoo! Finance Tax Center - File online. File on time.
> > > http://taxes.yahoo.com/filing.html
> > >
> > > ---------------------------------------------------------------------
> > > To unsubscribe, e-mail: users-unsubscribe at gridengine.sunsource.net
> > > For additional commands, e-mail: users-help at gridengine.sunsource.net
> >
> 
> ---------------------------------------------------------------------
> To unsubscribe, e-mail: users-unsubscribe at gridengine.sunsource.net
> For additional commands, e-mail: users-help at gridengine.sunsource.net
> 
> 


---------------------------------------------------------------------
To unsubscribe, e-mail: users-unsubscribe at gridengine.sunsource.net
For additional commands, e-mail: users-help at gridengine.sunsource.net




More information about the gridengine-users mailing list