An efficient and accurate metadataoriented prefetching scheme is critical for achieving the best metadata service performance in large distributed storage systems. Most previously developed algorithms, however, do not consider the characteristics of metadata, such as small size. In light of this issue, we presented a 3gram based model and datamining technique to use the traces of metadata accessing to prefetch groups of users’ future metadata operations. By using large tracedriven simulations in which our new prefetching scheme was adopted, it is shown that the hit rate for metadata access extracted from the traces can be increased by up to 3.9% and 16 % compared with NEXUS and least recently used(LRU), respectively.