HI, Adriam We have a similar implementation as youtobe . But when we deploy squid for large files, we found quick_abort does not work well. Problem looks like this: When cache miss , it is very hard for squid to get whole large file and cache it if there are lots of multi-thread download tools, like flashget or netransport are accessing this file at the same time. Observing through sniffer, we found the back side connection often stopped at the point of the first thread stopped. ->-----邮件原件----- ->发件人: Adrian Chadd [mailto:adrian@xxxxxxxxxxxxxxx] ->发送时间: 2006年9月2日 8:58 ->收件人: Mark Nottingham ->抄送: squid-users@xxxxxxxxxxxxxxx ->主题: Re: Large Files -> ->On Fri, Sep 01, 2006, Mark Nottingham wrote: ->> I'd appreciate some enlightenment as to how Squid handles large files ->> WRT memory and disk. ->> ->> In particular; ->> - squid.conf says that memory is used for "in-transit objects." ->> What exactly is kept in memory for in-transit objects; just metadata, ->> or the whole thing? -> ->Squid used to keep the whole thing in memory. It now: -> ->* can keep the whole object in memory ->* If memory is needed, and the object is being swapped out Squid will -> 'free' the start of the object in memory - the in-memory copy is -> then not used to serve replies but is just there to be written to -> disk. All subsequent hits come from disk. -> ->> - if something is in memory cache, does it get copied when it is ->> requested (because it is in-transit)? -> ->The whole object isn't copied during a memory hit; only the current ->'4k' range being read. -> ->> - How does sendfile support in 2.6 affect this? -> ->Sendfile support? :) -> ->> - Does anyone have any experiences they'd care to relate regarding ->> memory-caching very large objects? -> ->Not yet; but there's a project on my plate to evaluate this to ->begin caching p2p and youtube stuff a bit better.. -> -> -> ->Adrian ->
