More Effective Distributed ML via a Stale Synchronous Parallel Parameter Server