In G1, when the concurrent marking phase cannot keep up with the application, a Full GC will collect the entire heap. This Full GC currently takes place serially and, depending on heap size, can be very slow. It would be nice to parallelize the Full GC operation to make Full GCs less painful for customers.