Re: pthread update

[Marc Slemko <ma...@worldgate.com>]

On Sat, 18 Apr 1998, Dean Gaudet wrote:

> Ok, I'm successfully running zeusbench against it, 10 simultaneous
> clients, for 60 seconds without crashes.  Performance is slightly worse
> than a non-threaded apache (they're both right around 1200r/s on this p-ii
> 266Mhz, 128Mb SDRAM system, across loopback, 6k file, linux 2.0.34pre8). 
> I'm not sure why yet -- that's what I'll be working on next. 

How does the memory use for a large number of connections compare?

Does your setup let you bump up the simultaneous connections until 
the stock Apache falls over due to thrashing then try your threaded
one with the same setup to see if it lives?

I wouldn't be suprised if the current performance numbers you are seeing
are more related to zb and your test setup than the code.

Re: pthread update

[Dean Gaudet <dg...@arctic.org>]

Yeah I suspect my test setup a bit -- after all, with zb in the mix the
TLB still gets thrashed because context switches from apache to zb and
back have to switch the page tables.  The box is on 10baseT right now
though, so it's hopeless to do anything over the net. 

If I try 120 clients, the threaded server edges past the non-threaded
server, but only by 1 or 2%. 

My current plans, in about the order I'll do them, are: 

- loosen the locks in alloc.c

- get rid of as much tls() crud in http_main as possible -- right now each
of those costs an out of line function call to get the tls pointer.  Clean
up some other leftover cruft from the unix process model code.

- write mod_mmap_cache which caches mmap()s, shared between threads

- get a machine on a real network

- maybe consider testing some of the modules

Dean

On Sat, 18 Apr 1998, Marc Slemko wrote:

> On Sat, 18 Apr 1998, Dean Gaudet wrote:
> 
> > Ok, I'm successfully running zeusbench against it, 10 simultaneous
> > clients, for 60 seconds without crashes.  Performance is slightly worse
> > than a non-threaded apache (they're both right around 1200r/s on this p-ii
> > 266Mhz, 128Mb SDRAM system, across loopback, 6k file, linux 2.0.34pre8). 
> > I'm not sure why yet -- that's what I'll be working on next. 
> 
> How does the memory use for a large number of connections compare?
> 
> Does your setup let you bump up the simultaneous connections until 
> the stock Apache falls over due to thrashing then try your threaded
> one with the same setup to see if it lives?
> 
> I wouldn't be suprised if the current performance numbers you are seeing
> are more related to zb and your test setup than the code.
> 
>