'stanbul 2004 
gy for serial 
on model. In 
anges to each 
ML files and 
to kick of a 
r difficulty in 
‚ed easy third 
1d machines. 
b control and 
platform (no 
tomatic drive 
' came across 
since then in 
OM) 
> Component 
uses remote 
by allowing 
the client. It 
rectly over a 
The theory 
ved less than 
ed evaluating 
r to newer 
03). 
d be handled 
ned to enable 
Mability and 
caling, while 
'osoft, 2001). 
omplicated to 
lowever, it Is 
arch did not 
)plications. It 
ent that could 
Iso require a 
/are price for 
ut computing 
roxies were 
ocess. These 
im each node 
1e installation 
> file servers. 
be utilized by 
) that provide 
s Distributed 
cations from 
up that uses 
e will finally 
cts. 
  
International Archives of the Photogrammetry, Remote Sensing 
  
  
Submitting Machine 
  
  
Figure 1. Distribution Model 
3. SETUP 
3.1 Requirements 
The installation should be done on a network domain that has a 
valid domain controller to avoid authentication problems when 
mapping network drives. The different machines involved in 
Condor HPC solution are: 
Pool Manager: acts as the central Condor pool manager that 
matches the jobs with the available machines. This is the most 
important machine and shouldn't run jobs in a production 
environment since it needs to be running all the time. 
Submitting Machines: these are the machines that could 
submit jobs to the cluster. Another option is to use one machine 
as the gateway to the cluster and submit jobs only from that 
machine. The latter has the advantage of localizing the different 
run time option used to optimize the distribution. 
Computation Nodes: these are the workers that do the 
computation. These nodes could be dedicated rack mounted 
machines or any machine in the office that is connected to the 
network. Workstations could also be configured to be both 
submitting and computation nodes. 
File Server(s): these are machine that provide the disk storage. 
The shared data files need to be on a Windows Server family 
OS or use Samba from Unix machines since Windows 2000/XP 
will only allow up to three clients to map its shared drives. 
License Manager(s): machines used to provide the LGGM 
floating licenses to be used by the submitting and 
computational nodes. The opportunistic nature of GPro may 
sometimes result in more machines being available at one time 
than there are licenses. The Condor proxies of GPro correctly 
handle this by resubmitting the job until it gets a license and 
runs to completion. 
Accordingly before installing a GPro Condor pool a user will 
have to select the pool manager machine, machines that are 
going to be used to submit GPro jobs (this are most of the time 
operators machines) and computation nodes. It is possible to 
combine the Pool/File/License Manager to the same machine if 
it has the capability. 
  
and Spatial Information Sciences, Vol XXXV, Part B3. Istanbul 2004 
3.2 Setup 
From a users point of view, the work required to set up the HPC 
environment is minimal. Simply install Condor on the 
workstations and change the program names in the GPro 
preferences to their respective proxy equivalent. For example, 
the rectifier — "ADSRectifier.exe" is changed to 
"ADSRectifyCondorProxy.exe". According to their needs users 
could mix and match which processes to distribute and which 
processes to run locally. The following steps explain how to 
install a HPC cluster for GPro using Condor. 
Condor Setup: 
l. Download Condor from 
"httpz//www.cs.wisc.edu/condor/". 
2. First install Condor on a server that is going to be the 
Condor Pool Central Manager. 
3. Install Condor on all the computation nodes. 
* Dedicated computational nodes should be 
configured so that their 'Job Start Policy' is 
"Always Run Condor jobs". This defines a 
very responsive pool. 
e All other workstations, for example that 
operators work on, should be configured 
using the default start job after 15 minutes 
of low CPU activity policy. 
4. Test the condor installation using the examples 
distributed with Condor. 
GPro Setup: 
There is nothing special about the Condor pool used by GPro. 
The only assumption made by the installation is that all the 
nodes have access to shared file servers. 
I. Select a shared file system to be mapped by all the 
computation nodes. 
Install and configure the Leica licensing client on all 
workstations that run and submit jobs to the pool. 
3. Install the distributed version of GPro that sets up the 
required proxy executables and configuration files to 
submit jobs. 
4. Modify the default Condor job submission files to 
reflect the selected shared file system. 
5. Change the program names to the proxies in the 
general preferences of GPro on all the submitting 
workstations 
6. Start submitting jobs to the pool. 
t2 
For easy upgrades GPro should only be installed on the server 
and not on each of the computational nodes. A typical 
distributed run is shown in Figure 2 below.