Maintaining reference graphs in fully decentralized systems

The recent shift in processor development leads to massive parallelism in multiand many-core processors. So, together with the wide-spreading use of cluster technologies, this development leads to huge networks of cores and processors. The development of applications that exploit this increasing parallelism is a hard task, for which only a small fraction of programmers is prepared. One approach to support the implementation of parallel applications for distributed systems, besides others described in this thesis as well, is to offer a single system image (SSI). An SSI completely hides the underlying hardware complexity and distribution of data and threads. Thus, it can be programmed as if it is a symmetric multi-processor (SMP) system, while the SSI afterwards handles load balancing and the placement of data and threads dynamically at runtime. Hence, it allows the maintainer to seamlessly add and remove processing elements at runtime, without disturbing the execution of the running application. This approach seems to have faded from prominence in the recent years, as the low publication rate indicates. In this thesis, I argue that the SSI approach should be examined again, in the new context of the increased on-die parallelism as well as the spreading use of cluster technologies. As a starting point, this thesis describes one of the main requirements of a distributed system, and thus an SSI: the possibility to migrate objects between nodes, e. g. to place data and threads closer to each other, or use added resources and evade nodes, which are scheduled for shutdown. Together with such an object migration, the system must offer referencing entities a mechanism to locate and retrieve migrated objects. To allow the SSI to scale with very large systems, it requires a fully-decentralized mechanism to avoid bottlenecks and single point of failures. The development, analysis, and evaluation of such a decentralized location and retrieval mechanism are the main topics of this thesis. It describes various possible approaches, of which the two main mechanisms are the reactive and the proactive location update approach. The reactive location update approach updates outdated references upon access. To be able to do this, the approach uses proxy objects, which remain on the object’s previous location and forward all subsequent access messages to the object’s current location. The proactive location update approach updates all referencing locations immediately upon object migration. This thesis describes the design and analysis of these protocols and their prototypical implementation. Moreover, it evaluates different simulation runs with implicit and explicit object migrations. The chief insight of this evaluation is that the proactive location update approach decreases the object access latencies significantly, if a huge number of object migrations take place. Nevertheless, this fast object access comes with a significant management message overhead. The conclusion is that this overhead is only worthwhile as an optional add-on, for example, for objects where a fast access is crucial for the application performance. Beyond that, this thesis implements and evaluates an access path optimization mechanism for the reactive location update approach. This access path optimization propagates location update