Igor Grudenic

Load balancing MPI algorithm for high throughput applications

MPI (message passing interface) is a standard API (application programming interface) used to create distributed applications. Development of distributed MPI software is a challenging task in which many difficulties may arise. Common problems include load balancing, deadlocks and livelocks. In order to tackle complexity of parallel programming various methods are developed. In this paper we present a load balancing algorithm (LBA) that can be used instead of developing custom parallel application. The algorithm uses MPI transport mechanism for internal signaling and synchronization. The LBA is most suitable for the tasks that involve high and continuous data throughput. It performance will significantly degrade if used in complex computation topologies.

Modeling and verification of MPI based distributed software

Communication between processes in a distributed environment is implemented using either shared memory or message passing paradigm. The message passing paradigm is used more often due to the lesser hardware requirements. MPI is a standardized message passing API with several independent implementations. Specification and verification of distributed systems is generally a challenging task. In this paper we present a case study of specification and verification of MPI based software using abstract state machines (ASMs).

Job status prediction - catch them before they fail

Jobs in a computer cluster have several exit statuses caused by application properties, user and scheduler behavior. In this paper we analyze importance of job statuses and potential use of their prediction prior to job execution. Method for prediction of failed jobs based on Bayesian classifier is proposed and accuracy of the method is analyzed on several workloads. This method is integrated to the EASY algorithm adapted to prioritize jobs that are likely to fail. System performance for both failed jobs and the entire workload is analyzed.

Caching in parallel BDD package

Fast computation of Binary Decision Diagram (BDD) operations is essential for efficient symbolic model checking. In order to improve symbolic model checking we parallelized BDD operations on network of workstations (NOW). Sequential BDD packages have always been sensitive to cache design, which is even more true in the parallel environment. Distribution of cache to the NOW made larger cache possible, but it implicated problems with references in distributed environment and posed limitations on ordering of computation. In this paper we address these issues with our distributed cache architecture.