Performance and Scalability of Preconditioned Conjugate Gradient Methods on the CM-5