Modellierung des Online-Steerings von Gridjobs als Zugriffe auf verteilten gemeinsamen Speicher

In this work, four tasks for online steering are distinguished: Communication, data consistency, automated evaluation and optimization, and data access.

The focus of this work is on the data consistency task. For this task, a new model is presented, which views online steering as accesses to data in a virtual distributed shared memory. To ensure the consistency of the data, the intra-process condition and the inter-process condition are identified. Both define rules for the data exchange. Dependent on the requirements of the application, both conditions must be satisfied or only the intra-process condition must be satisfied.

Two consistency models are defined which satisfy the intra-process condition which are named special weak consistency and delayed weak consistency. Based on an algorithm from CUMULVS, the schedule consistency is defined which satisfies both conditions. Furthermore, data objects may exist for which none of the conditions must be fullfilled.

When using the special weak consistency or the delayed weak consistency, only pairs consisting of one application process and one steering process communicate. For both models, an update protocol and an invalidate protocol are developed and implemented. Additionally, an update protocol and an invalidate protocol for the PRAM consistency are implemented. With the implementations the behavior of the protocol was evaluated. Hereby, it was detected that the protocols of the special weak consistency were slower than the protocols of the delayed weak consistency. The performance difference increases if the round trip time in the network increases.

For online steering of Gridjobs, a possibility must exist to communicate with a Gridjob. In this work, a communication channel is developed which allow to establish a secure communication connection to a Gridjob without compromising the security of the site. Several scenarios are developed for establishing a communication channel, depending on the site's configuration. The appropriate scenario can be automatically detected.

The throughput and the round trip time of the communication channel are evaluated using different scenarios and security levels. The behavior of the communication channel can be modeled with a pipeline. The throughput is limited by the slowest pipeline stage. When using encryption, the throughput between German LCG-sites is limited by the computing time for encryption/decryption on the endpoints of the communication. When using two security levels without encryption, the bandwidth limits the throughput for one-directional data transfer. In the case of two-directional data transfer, the computing time at the endpoints can limit the throughput.

Every component adds a constant to the round trip time. Compared to the round trip time between sites, the additional contributions are negligible.

Optimization and data access are not discussed in the same level of detail than data consistency and communication. The presented optimization approaches are based on the assumption that the bandwidth are limiting the performance. The optimization selects the invalidate or update protocol based on predictions of their performance.

The goal of the data access task is, to intercept data accesses to steered date objects and call methods of the steering system in order to allow the steering system to react on data accesses. Hereby, classifications are made which discern the possibilities of the steering system to react on the data accesses.

With the implementation of the steering model in RMOST, jobs of the ATLAS experiment can be steered.