Re: General Runtime Information - Requests given to unlinkd

On Sun, 20 Sep 1998, Chris Wedgwood wrote:

> ... I'd like to
> see almost all IPC done via shared memory with more finely grained
> locking of objects where possible. Pipes may have to be used to
> wake-up sleeping process/threads.
>
> The reason I'd like to see these changes made it to get greater
> scalability on the networking side of things for large MP boxes,

Care to benchmark the performance of a Unix pipe compared to a shared memory
with locks? Any known studies that have done that?

I would be interested to see if there is a big difference for *Squid IPC
traffic patterns*. In case memory locks win big, somebody might want to patch
things to use [probably less-portable?] shared memory primitives. If not,
there will be no reason for extra coding.

Just a thought; before the list will be filled with long discussions on
shared memory pros and cons. :)

Alex.
Received on Tue Jul 29 2003 - 13:15:54 MDT