A system and method for efficiently removing duplicate data blocks at a fine-granularity from a storage array. A data storage subsystem supports multiple deduplication tables. Table entries in one deduplication table have the highest associated probability of being deduplicated. Table entries may move from one deduplication table to another as the probabilities change. Additionally, a table entry may be evicted from all deduplication tables if a corresponding estimated probability falls below a given threshold. The probabilities are based on attributes associated with a data component and attributes associated with a virtual address corresponding to a received storage access request. A strategy for searches of the multiple deduplication tables may also be determined by the attributes associated with a given storage access request.