For efficiency and portability across UNIX-like platforms, the cache implements its own non-blocking disk and network I/O abstractions directly atop a BSD select loop. The cache avoids forking except for misses to FTP URLs; we retrieve FTP URLs via an external process because the complexity of the protocol makes it difficult to fit into our select loop state machine. The cache implements its own DNS cache and, when the DNS cache misses, performs non-blocking DNS lookups (although without currently respecting DNS TTLs). As referenced bytes pour into the cache, these bytes are simultaneously forwarded to all sites that referenced the same object and are written to disk, using non-blocking I/O. The only way the cache will stall is if it takes a virtual memory page fault-and the cache avoids page faults by managing the size of its VM image (see Section 2.9). The cache employs non-preemptive, run-to-completion scheduling internally, so it has no need for file or data structure locking. However, to its clients, it appears multi-threaded.