Re: General Runtime Information - Requests given to unlinkd

On Sun, 20 Sep 1998, Chris Wedgwood wrote:

> On Sat, Sep 19, 1998 at 08:57:34PM -0600, Alex Rousskov wrote:
> > Care to benchmark the performance of a Unix pipe compared to a
> > shared memory with locks? Any known studies that have done that?
>
> Done right, shared memory with user-space locks typically take 10s of
> cycles or less.
> Using a pipe or fifo is going to take 100 times or longer.
> Not that either of these things are necessarily relevant to squid.

That is why I would like to see a convincing proof for the first two
statements via a benchmark that simulates *Squid IPC traffic patterns* first.

I understand that a single shared memory access might be cheaper than pipe
I/O. However, it is not clear to me if the difference is significant when
real traffic, context switches, etc are involved. Especially if a "highly
non-portable assembly code" may be added to Squid as a result of a change.

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