PhDOOS Logo

Supported by AITO.

These pages are currently maintained by Alex Potanin

 

Active objects management in distributed systems

Cyril Ray
( French )

IRENav - French Naval Academy
IRENav - Ecole Navale BP600 29240 Brest Naval
tel: 02 98 23 36 11
fax:
ray@ecole-navale.fr

Keywords:

concurrency distribution frameworks

Abstract:

Networks of workstations become a more and more widespread computing environment, in particular because of their attractive performance/costs ratio. It allows to distribute the resources on different computers in a transparent way for users and allows distributed computing with good performances. In this framework, we are interested in solving fundamental problems of distributed computing and specifically the management of huge set of objects (at least one million). The objects we consider are particular in the way that they have their own autonomy, they are active objects. More generally the studied topics are the object migration/replication and their consistency, the granularity of active objects, the paradigms of communication like objects based distributed shared memory without forgetting networks performance aspects on object communication.
In this research started in october 1999, we first set the problem of managing a huge amount of active objects in a single computer context and then proposed a distributed context as a first solution to overcome the limits of a sequential approach. These limits are mainly arising from the physical limits of today workstations. Data distribution on several computers enable us to increase the number of objects we can simultaneously run. However this solution which increases the concurrency between active objects is not completely satisfactory because the number of active objects is still too big for network architectures of 4 to 16 nodes.
We also started a study about differents communication paradigms. These paradigms are Remote Procedure Call (RPC), Message Passing (MP) and Distributed Shared Memory (DSM). The aim is to thoroughly study the properties and limits of communication middleware for active objects in distributed environments. In order to validate this study we are currently working on performance benchmarks for these different techniques. We already have tested a few products like Java RMI or Mico for RPC or more cautiously MPI and PVM for message passing. Concerning the distributed shared memory, we devellop an implementation of the K-managers protocol which defines a fail-safe object DSM.
Finaly, the last topic concerns migration and replication of objects. In the management of millions of objects, a good initial distribution isn't enough for optimal computation. It is then, necessary to migrate and replicate the objects from a node to another one in order to improve the global computation efficiency. We proposed a protocol, so-called Round Trip Bus protocol (RTB), that allows to control the load of each nodes of the distributed architecture and to improve the average communication time between objects thanks to a judicious migration.
We do want to actively contribute to the development of fundamental methods and tools for distributed object computing. We'd also like to consider the applied context for which our work could have significant impact. In this applied context the million of objects we talked about previously are nautical soundings. The management of a huge amount of objects is a delicate problem still not studied enough. It is difficult to keep a fine grained objects for users without being confronted to the limits of both computers and their operating systems. The evaluation of the different techniques of communication and the services they offer is an important step in the choice of development of middleware of communication which can support millions of objects. The network latencies are so strong that it is necessary to evaluate as better as possible the time taken by messages and the overhead resulting in the distributed computation. The RTB protocol enables to control dynamicaly the distribution of the millions of objects. For this protocol we are currently considering many improvements such as reliable communication or the treatment of migration urgency.

The PhD work started: October 1999

NO HTML3
NO JAVA

 

Last modified on Mon Aug 15 14:59:24 2005