Task-to-processor allocation for distributed heterogeneous applications on SMP clusters

Today, distributed architectures are based on multi core SMP nodes. Several middleware, like the FlowVR framework, exist to design interactive and heterogeneous distributed applications for these architectures. FlowVR defines an application as a set of codes mapped on cluster nodes and linked by a communication and synchronization network. But if we control the way modules are synchronized and mapped on nodes, we do not control the scheduling of concurrent modules on processors which is done by the Operating System scheduler. Since modules may be synchronized across different nodes, each local scheduling can affect the whole application performance. Moreover the OS scheduling is dynamic thus it is very difficult to predict and guarantee performance of such applications and especially of interactive ones because performance variations may totally break the immersion feeling. In this paper we propose to use a performance model to determine the processor load required by each distributed task. With this information we can define a task-to-processor allocation and tune the scheduler to respect it. Thus we can abstract this allocation from a particular OS, avoid effects of a dynamic scheduling, guarantee performance, optimize the processor use and the number of nodes used for our application.