Virtualized platforms on multicore systems are becoming increasingly common, especially since single-core performance is now increasing ever more slowly. Virtualized platforms, however, introduce additional cache-related nondeterminism to the equation, since additional cache misses may now occur because of the additional indirection between virtual central processing units (CPUs) and real CPUs. Analyzing the worst-case effects originating from such setups is paramount for real-time systems.

This paper contributes to the current toolset by adding the analysis of cache-related overheads to existing analysis techniques. It first presents the virtualization model, existing analysis techniques, and the modeling abstractions used in existing analysis techniques. Then the current so-called multiprocessor periodic resource (MPR) model is incrementally improved into a deterministic variant over the next two sections. This model is then the basis for presenting a method for computing necessary scheduling guarantees under the assumption that cache overhead is minimal. This is followed by the cache-related overhead scenarios that may occur in the virtualization scenario.

The next part of the paper describes analysis techniques in the presence of cache-related overheads. The first technique is task-centric, the second on the overall model. The hybrid technique finally shows how both techniques can be combined. The last section of this part describes the results of applying the methods on simulated random workloads, especially discussing the respective tradeoffs.

The paper ends with the usual discussion of related work and a wrap-up conclusion. It is very readable while still thorough and complete, even if you are not in the area of real-time programming. I found this paper quite relevant since it clearly discusses all the cache-related issues a system faces in a virtualized environment.