Web caches such as Squid don't usually come right out and tell you when disk I/O is becoming a bottleneck. Instead, response time and/or hit ratio degrade as load increases. The tricky thing is that response time and hit ratio may be changing for other reasons, such as increased network latency and changes in client request patterns.
Perhaps the best way to explore the performance limits of your cache is with a benchmark, such as Web Polygraph. The good thing about a benchmark is that you can fully control the environment and eliminate many unknowns. You can also repeat the same experiment with different cache configurations. Unfortunately, benchmarking often takes a lot of time and requires spare systems that aren't already being used.
If you have the resources to benchmark Squid, begin with a standard caching workload. As you increase the load, at some point you should see a significant increase in response time and/or a decrease in hit ratio. Once you observe this performance degradation, run the experiment again but with disk caching disabled. You can configure Squid never to cache any response (with the null storage scheme, see Section 8.7). Alternatively, you can configure the workload to have 100% uncachable responses. If the average response time is significantly better without caching, you can be relatively certain that disk I/O is a bottleneck at that level of throughput.
If you're like most people, you have neither the time nor resources to benchmark Squid. In this case, you can examine Squid's runtime statistics to look for disk I/O bottlenecks. The cache manager General Runtime Information page (see Chapter 14) gives you median response times for both cache hits and misses:
Median Service Times (seconds) 5 min 60 min: HTTP Requests (All): 0.39928 0.35832 Cache Misses: 0.42149 0.39928 Cache Hits: 0.12783 0.11465 Near Hits: 0.37825 0.39928 Not-Modified Replies: 0.07825 0.07409
For a healthy Squid cache, hits are significantly faster than misses. Your median hit response time should usually be 0.5 seconds or less. I strongly recommend that you use SNMP or another network monitoring tool to collect periodic measurements from your Squid caches (see Chapter 14). A significant (factor of two) increase in median hit response time is a good indication that you have a disk I/O bottleneck.
If you believe your production cache is suffering in this manner, you can test your theory with the same technique mentioned previously. Configure Squid not to cache any responses, thus avoiding all disk I/O. Then closely observe the cache miss response time. If it goes down, your theory is probably correct.
Once you've convinced yourself that disk throughput is limiting Squid's performance, you can try a number of things to improve it. Some of these require recompiling Squid, while others are relatively simple steps you can take to tune the Unix filesystems.