SUMMARY TO MY THESIS OF THE END OF MY STUDIES:
(SUPPORTED ON THE SUNDAY, JUNE,the 27th, 1999 in the computer science institute- university of sciences and technology houari boumediene algiers algeria, with the mention of"very good")
Title : putting on work of an atomic commitment protocol based on the consensus.
In a distributed system, a given application maybe brought to manipulate informations located on distinct sites. The putting on coherence of these informations is a complex problem which was the subject of many research works.
The tolerance of failures to conserve the coherence of databases, is among the problems which happens because of the distributed aspect of the distributed systems. Tolerate failures at the time of the atomic commitment of transaction returned to take a consistent decision even in the presence of failed sites.
The commitment operation of a transaction T, concens all the participants sites in the treatement, and consequently the data manager of the initiator site of T, has to pass the operation of commitment to all sites where is located data reached by T. what is true for the operation "Commit" remains true for the opeartion "Abort". Therefore, the treatement of a logical action (Commit or Abort) has to take place in many sites in the distributed databases manager system, and it's this which differ essentially ; the distributed treatement from the centralized one in the transactional systems.
The problem is more complicated in comparison to which it can appears in a first view. The sending of the "Commit" operation - by the manager of the distributed transactions - to all sites is deficient, because a transaction is not committed by the sending action, ( of the transactions manager ) but rather by the action of execution of the commitment (by the data manager). It's possible that the data manager send a "Commit" to the scheduler but this last reject it and consequently abort the transaction. In this case, if the database is distributed, the transaction has to be aborted in all sites which contains data used during its execution life. A local data manager, executes the final decision (Commit or Abort), which has been taken by the participant site at the end of the execution of an algorithm called atomic commitment protocol.
An atomic commitment protocol is a set of algorithms which allows to the participants to "Commit" or "Abort", a distributed transaction, and this with a common agreement.
In the thesis, we studied two types of protocols which resolves the problem of commitment in the distributed systems : The two phase commitment protocol, and a protocol of commitment based on the consensus. The first studied protocol is the protocol which is more known and more used (example: The CA-OpenIngres system use it in the distributed commitment [ZER 96]). The protocol allows to effect the commitment in two phases : a phase of vote and another of decision.
The major inconvenience of such protocol is that it's can lead to a total blocking state of system if only one site fails, which encouraged the research in the domain of algorithms tolerants of failures.
Among the paradigmes which allows to simplify the construction of applications tolerants of failures in the distributed systems, the problem of consensus is "without doubt" one of the most interesting, in the measure where it allows to abstract many problems of agreement. Every solution to the consensus problem can serve as a brick of base to resolve the problems of agreement [BAD 98]. The chapter 2 of thesis was consecrated for the study of the atomic commitment with the use of the consensus.
The solution of the consensus problem, used in the commitment is the one introduced by Chandra and Toueg with the use of the class of the failures detector à S [CHA 96a].
The use of consensus allowed to us to understand how a such paradigm allow to resolve the problem of non-frozen atomic commitment.
In fact, the obtained protocol is not frozen, despite of the presence of failures in the network, the protocol works and the operationnel sites ends their executions by taken one decision of consensus which is whether commitment or abortion of the distributed transaction.
The study of the two protocols of commitment, introduced in the thesis, allowed to us to know better the system programmation in a distributed environnement and to well master the working of distributed algorithms. This study allowed to us also to deepen our knowledges in the domain of exploitation of consensus problem, and more specifically its use in the applications tolerants of failures which still rest a subject of research.
As perspectives to our work, we propose a comparative study between the 2PC and the protocols based on the consensus by the use of classes à S and S, by an experimental study in a real network (with a large number of sites) and the evaluation of performances to base on our study that we realized, for the commitment of the results which we found.
Finally, we say that "the research in the domain of atmic commitment in a distributed systems and the exploitation of the paradigme of consensus" rest a domain which is widely open.
Remark : The software of the thesis is the language C under the UNIX system.
From the left to the right : ( Mr. REZGUI SAMI, Mr. LEMLOUMA TAYEB, Mr. HAMADI ALI, and a 4th colleague). an unforgettable photograph !!
A photograph which had been taken on the Sunday,June, the 27th, 1999 at the Computer Science Institute - University of Sciences and Technology Houari Boumediene ( U.S.T.H.B ) Bab-Ezzouar Algiers Algeria.
N.B : For further information about this thesis, or about its subject, would you contact me on: