In multi-channel architecture SSDs, as two or more flash memories can operate independently via different channels, I/O operations are actually performed in a cluster unit consisting of several pages. This paper proposes a novel write buffer replacement policy that fully exploits such high parallelism of SSDs and leads to better I/O performance. Our proposed scheme evaluates buffers in a cluster unit by intelligently aggregating the re-reference probability of pages in a cluster. The pages belonging to the least valuable cluster are evicted together when the replacement is needed. This replacement policy efficiently takes both the parallel architecture of SSDs and the temporal locality of I/O workloads into consideration, and thus achieves better I/O performance of SSDs. Through the trace-driven simulations, we show that our scheme improves performance up to 25% in terms of write throughput for common workloads.