Check out the new USENIX Web site. next up previous
Next: Limitations Up: Performance Effects Previous: Micro Benchmarks

Macro Benchmarks

The SPEC95 integer benchmark suite was also run to establish macro benchmarks (see Appendix 1 for environmental details). The results appear below in Figure 3.

Figure 3: SPECint95 benchmarks

The performance penalties measured by the SPEC95 integer benchmark suite indicated that StackGhost only shaved a few hundreths of a point off the speed metrics. The performance aberrations may be more attributable to noise than the StackGhost mechanism. An un-StackGhosted OpenBSD 2.8 kernel had slightly worse performance than a StackGhosted version in some instances, the only other explanation is that it is due to cache or TLB effects.

Discounting any noise in the benchmark, the geometric mean SPEC rating showed a StackGhost overhead of 0.1% with a Per-Kernel cookie and a 0.4% overhead with a random Per-Process cookie.