Tuesday, June 3, 2008

Modeling the Management Aspects of a Grid infrastructure

A common way to define the management aspects of Grid systems is the goal of the Reference Model Working Group. Paul Strong from eBay Research is involved in this area since long time and joined OGF thanks to the GGF/EGA merge. In the Reference Model WG session, Paul presented the status of the current model. They are close to publish an updated document for version 2.0 of the model. In Version 2.1, they plan to align with the GLUE Information Model. While the Reference Model is about the internal management aspects of a Grid infrastructure, the GLUE Information Model is about modeling the capabilities of a Grid service that are important to be exposed (see this document for more insight).

The typical question is "Why you are not using CIM?". Paul says: "CIM does not address aspects that are relevant to fast-changing e-infrastructures where the need for provision new services is fast. The notion of lifecycle is not present and also while the physical level is modeled in details, CIM lacks satisfactory models for the service level".

Two implementations are described, one from Fujitsu (presented by Michel Drescher) and one from eBay. Fujitsu focuses mainly on lifecycle aspect, while eBay focuses on relationships among components. Paul: "We built a 2-tier management application and we support semantic queries. For instance, if a user press the button X on the website, what is the physical component which will be involved?". "We also have a flash-based UI to present a mash-up of our management information. We roll out code every two weeks".

While plugging the monitor cable, Paul says "The great advantage of using OWL is that you can start with a simple model and then you can subclass later".

Michel shows a short demo about the Fujitsu implementation. Using a java GUI, he shows a few steps in inspecting the management aspects of a resource.

Then Paul shows the eBay interface: "The UI is dynamically generated by the underlying information. If the underlying components and relationships change, then we don't need to change the UI. We use standard SPARQL calls." Paul continues to show the aggregated resources for the search service in the eBay production infrastructure, 4.000 physical server in 4 locations (cool!). "I can ask: show me what this component is composed of" and a graph-like diagram appear with circles representing components and lines representing relationships. "At the moment, we see what was configured. We are working now on having real-time update on the ontology."

Paul: "We are working on enabling queries such as: show me the close components common to these two components which are performing bad; you may discover that there is a common network which could be the reason."

The final words on this very interesting session is that, given the complexity of the problem, only communities can build open source management software for managing e-infrastructure. It is very unlikely that vendors will support the heterogenous management space that you can find in systems like Grid. Leveraging the Semantic Web technologies in this area enable to model e-infrastructures by re-use the knowledge and extending core management models via iterations.

1 comment:

College Term Papers said...

It is an interesting and mind opening blog,I really didn't know about that, but you not only be informed me but I also getting knowledge from your article.