Should we remove the gbenchmark dependency from the mini runners?

[linmagit]

We implement the mini (operating unit) runners on top of the gbenchmark framework. We use the setup/teardown, benchmark filter, and the argument generation features of gbenchmark. But we're not using any metrics recording features of gbenchmark, nor we're reporting any "benchmarking" results. Should we remove the dependency to gbenchmark and just implement whatever we need? Does it worth the refactoring?

[17zhangw]

Just my thoughts on removing the gbenchmark framework. Refactoring out the dependency is not too difficult and would clean-up some messy parts of the mini-runners (code related to starting a particular runner). Removing gbenchmark might allow for more flexible argument generation, but we would need to implement an "output tracker" of sorts.

[17zhangw]

#1393 presents another reason for removing the gbenchmark dependency. Removing the dependency will allow more fine-grained control over how different runners are interleaved.