On Wed, 29 Jun 2005, Brett Glass wrote:
I've seen your messages describing long term plans for Squid 3.1, but we urgently need to solve problems with Squid and Windows Update quickly. We have clients using Squid whose Internet connections were literally taken down by the problems I mentioned in my message. And we are having severe problems trying to implement workarounds on our own servers.
Ok.
We've set the maximum object size on our main cache to 800 MB, but this causes Squid to cache many things we don't want to cache.
Like what?
It is also causing Squid run out of RAM and restart several times a day.
This should not happen. At least not with the current release (2.5.STABLE10 or later, before the 2GB patch). On earlier versions the situation you describe may result in excessive memory usage depending on timing of the requests and speed of the clients.
And we cannot afford not to cache Windows Update, because it will tax our bandwidth to the limit.
There is no easy solution short of disabling the range prefetching (range_offset_limit 0), and working on having partial object caching implemented for caching windows update or accepting that it's not cached by the normal traffic. At least I haven't been able to come up with any easy approaches. Implementing partial object caching requires some weeks of developer time which is a somewhat scarse resource in the current situation.
Only implementing the acl support in range_offset_limit requires less time, but still it requires probably a day or two depending on how skilled the C programmer looking at the problem is, maybe three if the programmer has not worked with Squid before.
As soon as you allow range prefetching you run into troubles due to "download accelerators" who greedingly fetches multiple concurrent ranges of the same object thinking it will speed up their download by slowing down everyone else. This unfortunately makes life very hard for caching proxies, and also steals bandwidth from others when there is congestion on the Internet.
Are there patches which can be applied to Squid 2.5 to fix it in the meantime? We and other users urgently need these, whether they realize what's slowing down their networks or not.
If there were easy patches, or even tricky ones for both allowing range prefetching and the use of download accelerators these would be available at least on devel.squid-cache.org. Today there is not. Sorry.
Regards Henrik
