We perform a characterization of traces in Java programs.  We collect traces 
in RedSpot, our trace collection system for the Jupiter Java Virtual Machine, 
using the SPECjvm98 and Java Grande benchmarks.  We characterize these traces 
according to their number, length, several forms of coverage, predictability, 
and execution patterns.  We find that traces can be used to group benchmarks 
into categories, represent a program compactly, focus optimization on only the 
most frequently executed instructions, and predict the control flow of a 
program.