Dynamic Profile of Instruction Sequences for the IBM System/370

Abstract

Instruction mixes such as the Gibson mix have been used for a long time as workload models for CPU's. However, since an instruction mix does not indicate the order of instruction execution, it is not suitable for the performance evaluation of advanced computers which employ pipelining. Instruction sequences are proposed as a generalization of instruction mixes. The instruction representation does not consist of a fixed number of instructions, but rather it is defined to be a string of consecutively executed instructions terminated by the instructions which potentially interrupt sequential instruction streams. Statistics on sequences have been collected from traces of problem state programs. It was found that there are not many distinct sequences and that they are short. The average sequence length may be less than seven instructions. A small subset of the distinct sequences, say less than 20 percent, accounts for more than 95 percent of all executed instructions.