CPU Resource Reservation for Simultaneous Multi-Thread Systems

Simultaneous Multi-Thread (SMT) techniques are becoming popular because they increase the efficiency of CPU resource usage by allowing multiple threads to run on a single physical processor at a very fine granularity. Emerging real-time applications, however, may not benefit from the SMT techniques because those techniques often compromise the predictable performance characteristics of applications, which real-time applications typically need to meet their computation deadlines. In this paper, we propose a new resource reservation scheme for SMT systems. In this scheme, a task scheduler dynamically enables and disables the SMT facility while real-time applications are running by monitoring the progress of the real-time applications. In this way, real-time applications can still meet their computation deadlines, and other best-effort applications can gain a high throughput due to the SMT facility. We have implemented this scheme on a Linux kernel and evaluated it on a Hyper-Threading processor, an Intel's implementation of SMT techniques. Our experimental results have shown that, for our workload, our scheme can guarantee real-time applications to use reserved resources while best-effort applications can obtain a high throughput due to the SMT facility.