Jyh-Herng Chow

A compiler framework for restructuring data declarations to enhance cache and TLB effectiveness

It has been observed that memory access performance can be improved by restructuring data declarations, using simple transformations such as array dimension padding and inter-array padding (array alignment) to reduce the number of misses in the cache and TLB (translation lookaside buffer). These transformations can be applied to both static and dynamic array variables. In this paper, we provide a padding algorithm for selecting appropriate padding amounts, which takes into account various cache and TLB effects collectively within a single framework. In addition to reducing the number of misses, we identify the importance of reducing the impact of cache miss jamming by spreading cache misses more uniformly across loop iterations. We translate undesirable cache and TLB behaviors into a set of constraints on padding amounts and propose a heuristic algorithm of polynomial time complexity to find the padding amounts to satisfy these constraints. The goal of the padding algorithm is to select padding amounts so that there are no set conflicts and no offset conflicts in the cache and TLB, for a given loop. In practice, this algorithm can efficiently find small padding amounts to satisfy these constraints.

False sharing elimination by selection of runtime scheduling parameters

False sharing can be a source of significant overhead on shared-memory multiprocessors. Several program restructuring techniques to reduce false sharing have been proposed in past work. In this paper, we propose an approach for elimination of false sharing based solely on selection of runtime schedule parameters for parallel loops. This approach leads to more portable code since only the schedule parameters need to be changed to target different multiprocessors. Also, the guarantee of elimination (rather than reduction) of false sharing in a parallel loop can significantly reduce the bookkeeping overhead in some memory consistency mechanisms. We present some preliminary experimental results for this approach.

On the linkage of dynamic tables in relational DBMSs

Tables and operations over tables are at the center of the relational model and have been at the core of the Structured Query Language (SQL) since its development in the 1970s. As database applications have grown rapidly, the concept of tables has been generalized in database languages. The new generalized table concept in the SQL standard and in some commercial databases includes explicitly defined derived tables, such as user-defined temporary tables, transition tables, user-defined table functions, and table locators, that can be manipulated by users. We call them dynamic tables, because their entities exist only at run time. The challenges that these dynamic tables pose to existing relational engines lie in the linkage between the creation of the derived table and its references. In this paper, we describe a uniform framework for compile-time and run-time processing of dynamic tables. We also give a thorough explanation of how such a generic framework can be realized in existing relational database management systems, such as IBM DATABASE 2TM Common Server. Our experience with our prototype has shown the simplicity, generality, and efficiency of our approach.