Re: Providing continuity of service during cache maintenance.

On Thu, 11 Dec 1997, Dancer wrote:

> What I wouldn't mind seeing is the network layer of Squid decoupled from the
> caching layer ...

Yes, I was also thinking about that. Here are some extra pro's:

Currently (1.18) all "ready" network and disk I/Os are handled in round robin
fashion inside one select() loop. This leads to very bad queuing delays when
the load is high (exponential growth with peak load). Looks like by isolating
two kinds of I/Os and implementing a simple scheduling to achieve fairness,
we can improve peak load performance quite a bit.

From the performance point of view, it may also be a good idea to have one
process/thread per every "device" (network card, disk driver or controller,
maybe even processor). This way Squid will benefit from multiprocessor
configurations and some nice scheduling/load control/balancing algorithms
become possible.

Squid 2.0? :)

Alex.
Received on Wed Dec 10 1997 - 20:02:27 MST