STAFF: A Model for Structure Layout Optimization

Abstract

Structure layout optimization has been proven effective in improving the performance of memory-intensive programs. Field-affinity-based models are widely adopted to guide optimization. It is both important and challenging to improve the models to yield higher optimization results. In response to this issue, this paper proposes STAFF, a new model to guide structure layout optimization. First, STAFF better models data field affinity using the spatial and temporal relations in affinity events (co-accesses of data fields). As a result, it can uniformly capture and distinguish various affinity relations in a fine-grained and time-sensitive manner. Second, it proposes a set of novel methods to estimate the contributions of affinity events with different spatial and temporal relations, and reweight them accordingly. Third, a method is proposed to jointly consider the effects of different data layout transformations, and the effects of the actual byte-arrangement and alignments. It supports a uniform impact estimation for structure layouts derived by applying combinations of structure splitting, structure peeling and field reordering in different sequences. The evaluation shows that, comparing with state-of-the-art methods, STAFF works up to 30% better, with a geometric mean improvement of 5%, demonstrating STAFF’s effectiveness.