On 12/12/2013 06:50 PM, Nathan Hoad wrote: > This leads me to believe that the objects that are consuming all of > the memory are genuinely using that memory, and are freed on shutdown. This hypothesis is relatively easy to test, especially if you can reproduce the high memory usage in a lab (or if you can control live load on a given Squid instance): Load Squid for a while. Then, instead of stopping Squid, let it sit for a few minutes without traffic so that there are no HTTP transactions waiting for timeouts and such. Then again load Squid with traffic for a few minutes. Repeat a few times while measuring memory usage. If, after a few iterations, the base of the memory usage humps stay about the same, then your theory is probably correct. If the base of the humps keep climbing up, there is an effective leak, even if all that accumulated memory is freed by Squid during a proper shutdown. The next step would depend on the above outcome. Either we will need to compare memory usage of v3.2 and v3.3 (to identify the objects or areas that started consuming more RAM) OR we can help you find the v3.3 leak (e.g., using valgrind or log analysis). HTH, Alex.