Due to the emergence of the HTTP/1.1 standards, persistent connections are increasingly being used in web retrieval. This paper studies the caching performance of web clusters under persistent connections, focusing on the difference between session-grained and request-grained allocation strategies adopted by the web switch. It is shown that the content-based algorithm considerably improves caching performance over the content-blind algorithm at the request-grained level. However, most of the performance gain is offset by the allocation dependency that arises when the content-based algorithm is used at the session-grained level. The performance loss increases with cluster size and connection holding time. An optimization problem is formulated to investigate the best achievable caching performance under session-grained allocation. Based on a heuristic approach, a session-affinity aware algorithm is presented that makes use of the correlation between the requests in a session. Experimental results show that while the session-affinity aware algorithm outperforms the content-based algorithm under session-grained allocation, this optimization cannot fully compensate for the performance loss caused by allocation dependency.