> On Mon, Sep 24, 2001 at 11:00:31AM -0400, Bill Stoddard wrote: > > >From what I have seen, pthread accept serialization does not scale well on > > >multi CPU > > Solaris systems (above 4 way). pthread is much faster than fcntl on single > > CPU boxes, but > > much slower than fcntl on a 12 way box. > > On Ian's 8-way box, pthread was faster - causing a ~60rps improvement > in httpd just from switching the lock type via AcceptMutex. >
Just occured to me that our measurements were on 1.3. If you are seeing better performance with pthread on an 8 way in threaded 2.0, then go for it. I do agree with jeff that we should -not- adopt pthread by default on all platforms. > If you guys have access to a 12-way box, could you grab the current > code and see if 12 CPUs is still much slower? Ian's seems to be > the current high end of what anyone has access to. If no one > has access to a 12-way box and we know it is noticably better > at least up to 8 CPUs, I have no problem justifying that is the > default configuration on Solaris. > > fcntl() on a 2-way Sol8 box here caused the httpd children to exit > at high load - I'm not sure if Ian saw this or not. Have you guys > seen such a situation? Havenot seen this in 1.3. Haven't tested in 2.0. > I detailed what I saw in STATUS. I'm leery > of seeing the default cause httpd to exit at high loads. -- justin >
