Examination of Real-Time Optimization Methods for High Performance Computing

Background

In our ongoing research project Real-Time Computing on Multicore Computers, we have developed a proof-of-concept for the execution of hard real-time code on one core of a multicore x86 processor. The main idea is to shield a process from task switches, interrupts and unpredictable system call latencies with as least modifications to the operating system as possible. While developed with hard real-time capabilities in mind, the methods might also be of benefit for high-performance computing.

With reduced Operating System interruptions, the cache can be used more efficiently. Furthermore, parallel applications suffer from delay in individual cores when the others must wait (OS jitter). This can be addressed by synchronizing the OS maintenance work on all cores.

In the scope of this work is to evaluate which real-time methods can be used with MPI and OpenMP programs and to evaluate the performance. To address the OS jitter, algorithms in user-space and kernel as well as kernel configurations should be developed and evaluated.

Assingment

• Evaluation of the impact on MPI applications
• Evaluation of the impact on OpenMP applications
• Performance Evaluation
• Analysis of Operating System Jitter
• Development of Methods to reduce OS jitter

We offer

• Comfortable working-environment, as well as a good equipped Computer infrastructure
• Personal and engaged supervision

Profil of requirements

• Intermediate diploma of Electro- and Informationstechnics or (technical) computer science
• Alternatively the according prerequisites for a Master's thesis
• Good programming skills in C/C++
• Understanding of operating system internals and system programming
• Linux experience recommended
• Competence to work autonomous and creative

Contact