using the same seed for all runs of a build will make it possible (easier) to compare results from different test runs (e.g. on different platforms, w/ different flags) and consequently will make test results analysis easier. the proposed solution is to use the seed based on Runtime.version() / "java.vm.version", which are different from build to build, if there is no seed specified by "jdk.test.lib.random.seed" property.

a version-based seed is used only for promotable builds (i.e. ones that have build number and it's greater than 0); local and remote/mach5 ad-hoc builds (by default) don't specify a build number, so a random seed value will be used for them. and as before, if "jdk.test.lib.random.seed" is set, its value will be used as seed oblivious to build type.