PRADA: Predictable Allocations by Deferred Actions

Abstract

Modern hard real-time systems still employ static memory management. However, dynamic storage allocation (DSA) can improve the flexibility and readability of programs as well as drastically shorten their development times. But allocators introduce unpredictability that makes deriving tight bounds on an application's worst-case execution time even more challenging. Especially their statically unpredictable influence on the cache, paired with zero knowledge about the cache
set mapping of dynamically allocated objects leads to prohibitively large overestimations of execution times when dynamic memory allocation is employed. Recently, a cache-aware memory allocator, called CAMA, was proposed that gives strong guarantees about its cache influence and the cache set mapping of allocated objects. CAMA itself is rather complex due to its cache-aware implementations of split and merge operations. This paper proposes PRADA, a lighter but less general dynamic memory allocator with equally strong guarantees about its influence on the cache. We compare the memory consumption of
PRADA and CAMA for a small set of real-time applications as well as synthetical (de-) allocation sequences to investigate whether a simpler approach to cache awareness is still sufficient for the
current generation of real-time applications.