fellowship position

[Holger Bischof]

Dear colleagues,

please inform potential candidates about
the following fellowship position:

http://www.coregrid.net/mambo/component/option,com_jobline/Itemid,93/task,view/id,1/

Thanks and regards,

Holger

Duration: 18 months
Hosts: University of Muenster, Germany
Technical University of Delft, The Netherlands

- Context

The fellow will work together with two research groups involved in
CoreGrid : the group led by Prof. Sergei Gorlatch at the University of
Muenster, and the group led by Prof. Dick Epema at the Technical
University of Delft (The Netherlands).

The group in Muenster is active in several research areas of parallel
and distributed computing in general and grid computing in
particular. In the area of component-based programming we focus on
efficient parallelization strategies for distributed machines spread
across wide-area networks and the development of application case
studies for such systems. To cope with the special characteristics of
computational grids, such as heterogeneity of employed machines and
programming languages, dynamic changes of the runtime environment etc.,
we introduced a novel component-based programming model: Higher Order
Components (HOCs) whose main advantages are ease-of-use and
efficiency. For implementing higher-order components we defined and
implemented HOC-SA, the service architecture for HOCs. HOC-SA includes a
mechanism for maintaining mobile codes using multiple communication
technologies. These technologies include SOAP, which is the standard
web services protocol but does not define any means for code mobility
itself. HOC-SA allows that the application developer only deals with
HOCs. The HOCs can be used similarly to higher-order functions in
functional programming and therefore only require application-specific
knowledge. All technical requirements such as efficiently exploiting
middleware systems (e.g., the Globus Toolkit), databases, parallel
processors etc. are handled by the HOC-SA. The recent implementation
includes a web portal, supports the most recent internet standards
(WSRF, WSN, etc.) and is tested using example applications (e.g. image
processing, mathematics, bioinformatics).

In the area of scheduling for grids, the group in Muenster has
experience in the prediction of execution times for grid components
running on high-performance grid servers. The group is investigating
methods for predicting the performance of complex programs by combining
prediction and monitoring information from different sources
(e.g. resource availability data, benchmark results, etc.) with
structural information for the program (e.g. workflow description). The
group also concentrates on methods and algorithms for static scheduling
of workflows based on predicted performance, and dynamic rescheduling of
workflows based on monitoring information and performance results
obtained from already executed jobs. The group in Delft is active in
the areas of resource management and scheduling, in particular in grids.

Our main goal is to design, implement, and analyze mechanisms and
policies for scheduling jobs in grids, and in particular, to allow jobs
to require co-allocation, that is, to require resources in different
subsystems making up a grid that are under the control of separate,
autonomous resource managers. The basic premise of this research is that
in the absence of co-allocation, a grid scheduler is nothing but a load
balancer which selects a single site for every job to run. Of course,
even when co-allocation is allowed, a grid scheduler will still act as a
load balancer for those jobs that cannot, or do not want to, employ
co-allocation. Over the last five years, we have performed extensive
simulations of different models for co-allocation. Such models can have
very many parameters, such as the number and sizes (in terms of numbers
of processors) of job components, the size of the Fellowship Programme
(FP) grid, the service-time distributions, whether there is only a
single or multiple schedulers in the system, the slow-down due to the
relatively slow wide-area communication, etc. Overall, we found that for
simple parallel applications, co-allocation may be beneficial for the
utilization of the system and the reponse times of the jobs.

Starting about two years ago, we are building an actual co-allocating
scheduler in our fivecluster, 400-processor DAS testbed, which in
addition to the required numbers of processors also takes into account
the locations of (large) input files. Our current prototype, which is
implemented with Globus Toolkit components, is very reliable, but one of
its most notable shortcomings is that it currently only accepts
straightforward parallel applications. One of our most important current
aims is to make it suitable for more complicated jobs, for instance,
workflows and component-based applications.

Qualifications: Postgraduate students, candidates must be fluent in
English an institute is not allowed to host a fellow of the same
nationality or a fellow who was previously affiliated to this institute.

Compensation Allowance: 2300 euros net per month
How to Apply Please use the online application form at
http://www.coregrid.net/mambo/content/view/98/104/
Please see detailled guidelines at
http://www.coregrid.net/mambo/content/view/92/98/

--
Holger Bischof | Tel.: +49 (0)251 83-32744
Universität Münster | Fax.: +49 (0)251 83-32742
Institut für Informatik | hbis...@uni-muenster.de
Einsteinstr. 62 | http://wwwmath.uni-muenster.de/pvs
D-48149 Münster, Germany |

--