چكيده به لاتين
Running of the next-generation of tightly-coupled high performance computing (HPC) applications faces the scalability challenge. The inherent dynamicity of tightly-coupled HPC applications, lacking knowledge about application’s workload and the complex computation synchronization behavior between their execution units, are major factors decreasing scalability of these applications. Partitioning of applications between groups of machines, based on static scheduling, and composition of tightly-coupled applications in each machine, are proposed in order to decrease the scalability degradation. The composition of heterogeneous HPC application on common physical machines requires virtualization and computational clusters based on virtual machines that cause scalability being negatively impacted by virtualization overheads and interferences. Dynamic management of workload and unknown runtime behavior of applications need dynamic scheduling. This thesis proves the possibility of scalability guarantee using scheduling for computational clusters based on virtual machines. To guarantee scalability, a framework called Taraz is presented that schedules execution units by considering effects of dynamic workload, heterogeneous system software, virtualization and idling due to the synchronization of execution units of applications, in the form of two parameters (1) the cost of computation idling due to synchronization, and (2) the cost of workload of execution units. Taraz uses two techniques (1) unification of idling and workload costs, and (2) creation of two abstraction levels process and virtual cluster spaces. In order to guarantee scalability, Taraz schedules execution units and processes simultaneously by assignment and reassignment procedures. In order to reassign execution units and process, Taraz defines a stability condition that its fulfillment makes the high performance virtual cluster stable. In this thesis, we prove that if Taraz is stable the scalability for the dynamic tightly-coupled HPC applications is guaranteed in high performance virtual vlusters.
