The performance and scalability of centralized network storage systems cannot follow the processing needs of rapid growing data. Using large-capacity client caches is one of the solutions. This paper points out that the efficiency of "data exchange" between phases is the most important problem in managing large-capacity cache. And then, it defines Stable Set Model (SSM), which can characterize the phase-transition behaviour of a data access stream, and obtain granularity selection and two localities, so as to increase the efficiency of client caches in "data exchange". Using SSM based methods can optimize the backend load and response time of network storage systems. Experimental results show that they can decrease backend load to 2.0%~15.8% and total response time to 0.8%~15.2% compared with traditional ones when cache capacity is large.