Hendrik Voigtl�nder wrote: > I am just not sure how to read those numbers. When should be a service > time considered to high? I am quite sure that my hit service time is > good, but I wasn't sure about the miss time.
Anything under ~ 1 seconds is probably fine for misses, and even up to 2 seconds depending on congestion and latency on your link. Remember, a cache miss requires at least 1 TCP connection to a remote server and possibly a DNS lookup (if it isn't cached). On top of that, there is the time to download the data itself and the latency of your connection. Additionally, a cache miss of a large file will take a long time to download, and many of these over time will skew the average. > Do you think there is any way for improving the hit ratios? Line is not > paid per month and not by traffic, but higher hit ratio should IMHO mean > further improving the service time. Look at the refresh_pattern setting in your squid.conf. You will probably get the most benefit by tuning settings for image files, which tend to be large, downloaded often, and seldom changed. >> Why are you concerned about performance? What bottlenecks are you seeing? > Main reason for asking: I am just puzzled, that 1000 Clients are not > pulling more than 40 requests/second @ 250kbytes/second. (We are talking > about a company here, not internet access point or similar.) The ufs store method is probably a major (if not the only) bottleneck here. Try to solve the problems you were having with diskd - switching to it will give you a performance boost over ufs. diskd has its own FAQ section with configuration tips - maybe it will help: http://www.squid-cache.org/Doc/FAQ/FAQ-22.html > pppoe on openBSD is not the fastest of his kind, probably there is room > for improvement. I may try to switch over to debian/linux on this > machine as well. If you switch to Linux, use aufs instead of diskd (it performs better). > I can not use delay pools, as I need to set smaller delays when going > direct over a much smaller line (when parent is down). That is why > parent connections are set to no-delay. I can not limit download size > either. I am really stuck here. How do you make that switch when the parent goes down? Adam
