This paper presents an empirical comparison of four different techniques for filtering large test suites: test suite minimization, prioritization by additional coverage, cluster filtering with one-per-cluster sampling, and failure pursuit sampling. The first two techniques are based on selecting subsets that maximize code coverage as quickly as possible, while the latter two are based on analyzing the distribution of the tests' execution profiles. These techniques were compared with data sets obtained from three large subject programs: the GCC, Jikes, and javac compilers. The results indicate that distribution-based techniques can be as efficient or more efficient for revealing defects than coverage-based techniques, but that the two kinds of techniques are also complementary in the sense that they find different defects. Accordingly, some simple combinations of these techniques were evaluated for use in test case prioritization. The results indicate that these techniques can create more efficient prioritizations than those generated using prioritization by additional coverage.