On 11/07/2017 04:42 AM, Vieri wrote: > So I'm worried that 32768 may not be enough. > Is this weird, or should I really increase this value? Think about the underlying physics of what you are observing. It may help reduce guessing and guide you towards a solution: You can estimate the reasonable number of file descriptors using expected maximum request rate and mean response time. Add, say 30%, to account for long-lived persistent connections. Remember that Squid uses one descriptor for the from-client connection and one descriptor for the to-server connection. If that estimate is way below 32K, then the current limit itself is not the problem. Otherwise, it probably is (and then you probably need more Squid workers not more descriptors per worker). * It is possible, perhaps even likely, that some some unknown problem suddenly consumes almost all CPU cycles, drastically slowing Squid down, and quickly consuming all file descriptors (because accepted connections are not served fast enough). * It is also possible, albeit less likely, that some unknown problem slows Squid down over time and slowly leads to excessive file descriptor use and even 100% CPU usage. To distinguish the two cases, consider studying transaction response times and the total number of connections logged every minute of every hour. You should collect such stats for many other reasons anyway! Alex. P.S. I trust you have already checked all system logs for more clues and found nothing of interest there. _______________________________________________ squid-users mailing list squid-users@xxxxxxxxxxxxxxxxxxxxx http://lists.squid-cache.org/listinfo/squid-users
