A mathematical model and a metaheuristic approach for a memory allocation problem

Abstract

International audienceMemory allocation in embedded systems is one of the main challenges that electronic designers have to face. This part, rather difficult to handle is often left to the compiler with which automatic rules are applied. Nevertheless, an optimal allocation of data to memory banks may lead to great savings in terms of running time and energy consumption. This paper introduces an exact approach and a vns-based metaheuristic for addressing a memory allocation problem. Numerical experiments have been conducted on real instances from the electronic community and on dimacs instances expanded for our specific problem

A new variable neighborhood search approach for solving dynamic memory allocation problem

“…We assume that moving data structure is allowed only in between two time blocks and not during, without loss of generality. The problem with only one time block is a static memory allocation problem and more about it can be found in [8] and [12]. Moving data structures and rearranging their allocation is useful because in some cases moving data structure from external to internal memory bank and loading it from the internal memory bank can be less expensive than loading and operating with it while it is mapped to the external memory bank.…”

“…in [14]. In [12] Soto et al proposed the first Mixed Integer Linear Programing (MILP) formulation for the static version of the memory allocation problem and a metaheuristic approach based on the Variable Neighborhood Search (VNS). Later, Soto et al [13] dealt with a dynamic version of the memory allocation problem, providing ILP formulation, and two iterative approaches for solving DMAP in embedded systems.…”

Dynamic Memory Management for Embedded Systems

“…It is possible to consider situations where the processor has simultaneous access to several independent buses; however, a greedy algorithm would not be enough for that new problem and a more complex approach such as the one presented by [151] would be needed. On the contrary, our algorithm can consider the effects of mapping DDTs with interleaved accesses into multi-bank DRAMs to minimize the number of row changes in the DRAM banks.…”

“…A hardware structure to transparently map stack data into a scratchpad memory is presented by [70]. In [151], the authors explore, both from the perspective of an exact solver and using heuristics, the problem of placing (static) data structures in a memory subsystem where several memories can be accessed in parallel. Their work is partially a generalization of the mapping phase in our methodology, if each of our pools is considered as a big static data structure (an array) with a fixed size.…”