Data deduplication is an effective method to reduce data storage requirements. In data deduplication process, fingerprint identification may cause frequent on-disk fingerprint lookups which hurt performance seriously. Some locality-aware approaches were proposed to tackle this issue. Recently, the Persistent Memory (PM) brings low latency and high bandwidth, and has become a hotspot in data storage. Deduplication systems with fingerprints stored on PM will provide extremely fast on-disk fingerprint lookup, and therefore traditional locality-aware approaches designed for slow devices are likely no longer valid.In this paper, we model the traditional locality-aware approaches and analyze their performance on PM. Inspired by the analysis, we propose an optimized PM-based fingerprint identification scheme in which the fingerprint cache is replaced with a simple, low-cost read buffer, and the order of the Bloom filter and the read buffer is swapped. The experimental results on real PM devices show that, compared with the traditional locality-aware approaches, the proposed scheme improves the fingerprint identification throughput by 1.2–2.3 times.