Multi-tier platforms are hierarchically organized multicomputers with multiprocessor nodes. Compared with previous-generation single-tier systems based on uniprocessor nodes, they present a more complex array of performance tradeoffs. We describe performance programming techniques targeted to finite difference methods running on two large scale multi-tier computers manufactured by IBM: NPACI’s Blue Horizon and ASCI Blue-Pacific Combined Technology Refresh. Our techniques resulted in performance improvements ranging from 10% to 17% over a traditional single-tier SPMD implementation.