Performance Implications of Multiple Pointer Sizes

An increase in address size is a discontinuity: the Abstract fraction of ‘‘interesting’’ bits in a pointer shrinks a lot. Programs that were pushing the limits of the old Many users need 64-bit architectures: 32-bit sysaddress space consume almost none of the new adtems cannot support the largest applications, and 64dress space. bit systems perform better for some applications. However, performance on some other applications While some programs will soon grow to need the can suffer from the use of large pointers; large new address space, many (if not most) will not. A pointers can also constrain feasible problem size. large fraction of the storage used for pointers often Such applications are best served by a 64-bit machine goes to waste. A program that needs 33 bits of adthat supports the use of both 32-bit and 64-bit pointer dress space will not run at all on a 32-bit system, but variables. a program that runs happily on a 32-bit system will ‘‘waste’’ more than half of the bits in every 64-bit This paper analyzes several programs and propointer. gramming techniques to understand the performance implications of different pointer sizes. Many (but not Does this hurt performance? In this paper, we all) programs show small but definite performance examine a number of ways in which pointer size afconsequences, primarily due to cache and paging effects performance for programs that do not need adfects. dresses larger than 32 bits. We will show that the effects depend on many variables, including problem size, intensity of pointer use, memory system design,