1. German Astrophysical Virtual Observatory
Gadget on GAVOGRID
AEI-Clustertag, GAVO and GridComputing

2. GAVO Architecture VO-Sites Grid Enabled Sites GAVO WebPortal
HTTP, SOAP
Secure Protocols
WebServices
GridServices
Temporary
DataSpace
UserSpace
ConeSearch
Matcher
Abstract View of current architecture:
GAVO uses two tiers of services:
Access to ConesSearch, Matcher through
http and soap-protocols, the services are running within
the normal levels of security (provided through the firewalls of the sites,
allowing acccess to dedicated machines on known ports
The Cone-Search queries archives in local databases and files and
VO-archives through http and soap-protocols
2. Access to Grid-managed ressources:
The matcher-engine validates itself to the Grid, to get access to computing
facilities for parallel-processing and temporary storage
Users, who obtained Login, get access to
Grid-managed ressources (parallel-computing jobs, runtime information, job-results)
Stores virtual data from matcher results or cone-searches
in userspace
Files DB
Computing Tasks
Grid security realm
Normal security realm
AEI-Clustertag, GAVO and GridComputing

3. GAVO-GRID (testbed) using Globus GAVO-Grid
AEI-Clustertag, GAVO and GridComputing

4. running MPI-application across site-boundaries:
GAVO-GRID
Testcase for GAVO:
running MPI-application across
site-boundaries:
Gadget is a widely used particle-code for
cosmological simulations
Gadget (MPI-version) requires 2n processors
Gadget roughly uses for each timestep
80% calculation,
20% communication
(GAlaxies with Dark matter and Gas intEracT )
AEI-Clustertag, GAVO and GridComputing

5. Running MPI-applications with Globus / mpichG2
GAVO-GRID
Running MPI-applications with Globus / mpichG2
- all nodes communicate with each other
(ip-adresses known)
- using ssh instead of rsh
- requires a range of TCP/IP-ports open
(globus-tcp-range)
- allocation of bandwidth
(GAlaxies with Dark matter and Gas intEracT )
AEI-Clustertag, GAVO and GridComputing

6. Results from tests with uniform nodes 167,05 0,11
GAVO-GRID
Results from tests with
uniform nodes
Wall time
Comm time
On board
167,05
0,11
Across site
179,69
1,05
(np = 8)
105,3
1,30
(GAlaxies with Dark matter and Gas intEracT )
AEI-Clustertag, GAVO and GridComputing

7. Results from tests with heterogenous nodes 162,03 0,11
GAVO-GRID
Results from tests with heterogenous nodes Np
Wall time
Comm time
On site
2/2 AMD/ XEON
162,03
0,11
Across site
4 AMD
97,7
1,16
4/4 AMD/
XEON
85,9
0,15
105,33
1,18
8/8 AMD/
77,63
1,96
(GAlaxies with Dark matter and Gas intEracT )
AEI-Clustertag, GAVO and GridComputing

8. (GAlaxies with Dark matter and Gas intEracT )
GRID-Collaboration
Simulations for cosmological problems improve with
the number of particles.
There is a clear relation
between available computing ressources and the
number of particles a simulation can handle.
Integration of clusters using grid-technology is an option
which can significantly lower the costs of getting bigger
simulations running.
One of the main issues then is allocation of
more bandwidth
(GAlaxies with Dark matter and Gas intEracT )
AEI-Clustertag, GAVO and GridComputing

9. GRID-Collaboration in Potsdam:
Potsdam has a good net infrastructure,
where more bandwidth is relatively easy
available
For testing purposes, a low cost solution
providing more bandwith
could be installed, including
AEI, AIP, PIK, Uni Potsdam
This could be a first step of building part of
the D-Grid at Potsdam.
(GAlaxies with Dark matter and Gas intEracT )
AEI-Clustertag, GAVO and GridComputing

10. are using grid-technology do demanding cosmological simulations
GRID-Collaboration
AEI and AIP
are using grid-technology
do demanding cosmological simulations
working towards a grid-integration of
their respective clusters
working on specific problems (visualization
of simulation results, distributed file systems)
(GAlaxies with Dark matter and Gas intEracT )
AEI-Clustertag, GAVO and GridComputing