We study replacement algorithms for non-uniform access caches that are used in distributed storage systems. Considering access latencies as major costs of data management in such a system, we show that the total cost of any replacement algorithm is bounded by the total costs of evicted blocks plus the total cost of the optimal off-line algorithm (OPT). We propose two off-line heuristics: MIN-d and MIN-cod, as well as an on-line algorithm: HD-cod, which can be run efficiently and perform well at the same time. Our simulation results with Storage Performance Council (SPC)’s storage server traces show that: (1) for off-line workloads, MIN-cod performs as well as OPT in some cases, all is at most three times worse in all test case; (2) for on-line workloads, HD-cod performs closely to the best algorithms in all cases, and is the single algorithm that performs well in all test cases, including the optimal on-line algorithm (Landlord). Our study suggests that the essential issue to be c...