You are viewing a plain text version of this content. The canonical link for it is here.
Posted to dev@httpd.apache.org by Dean Gaudet <dg...@arctic.org> on 1998/07/06 04:35:07 UTC
record specweb...
http://www.sun.com/solaris/xeon/
I'm still having a hard time convincing myself it's not cheating to put
http into the kernel. ;)
Dean
Re: record specweb...
Posted by Dean Gaudet <dg...@arctic.org>.
That's sort of what the flow stuff is all about... providing an API for
modules to give enough information so that a cache can be built which
supports all sorts of applications. static stuff is the easiest.
Dean
On Mon, 6 Jul 1998, Ian Kallen wrote:
>
> FWIW, I really like the idea of having the listening and translation and
> all that stuff built into the FreeBSD kernel though I've no idea how y'all
> would graft the API stuff in - I wouldn't want to sacrifice all of wild
> and wooly authentication, mod_perl, mod_rewrite and other extensibility
> stuff just to get raw request rates up. I'll leave it to you geniuses to
> figure that out :)
> Ian (it's an Apache world, we just live in it) Kallen
>
> On Mon, 6 Jul 1998, Dean Gaudet wrote:
> :No, you really need to get into the kernel to get that high. flow gets
> :some of the way there, but to go all the way you need a kernel module (you
> :need flow in the kernel ;).
>
> --
> Peace was the way.
> -- Kirk, "The City on the Edge of Forever", stardate unknown
>
>
Re: record specweb...
Posted by Ian Kallen <ia...@gamespot.com>.
FWIW, I really like the idea of having the listening and translation and
all that stuff built into the FreeBSD kernel though I've no idea how y'all
would graft the API stuff in - I wouldn't want to sacrifice all of wild
and wooly authentication, mod_perl, mod_rewrite and other extensibility
stuff just to get raw request rates up. I'll leave it to you geniuses to
figure that out :)
Ian (it's an Apache world, we just live in it) Kallen
On Mon, 6 Jul 1998, Dean Gaudet wrote:
:No, you really need to get into the kernel to get that high. flow gets
:some of the way there, but to go all the way you need a kernel module (you
:need flow in the kernel ;).
--
Peace was the way.
-- Kirk, "The City on the Edge of Forever", stardate unknown
Re: record specweb...
Posted by Dean Gaudet <dg...@arctic.org>.
On Mon, 6 Jul 1998, Christof Damian wrote:
> Dean Gaudet wrote:
> >
> > http://www.sun.com/solaris/xeon/
> >
> > I'm still having a hard time convincing myself it's not cheating to put
> > http into the kernel. ;)
>
> ;-)
>
> I wonder how high one of these new four way Xeon boxes (Dell PowerEdge
> 6300, IBM Netfinity 7000 M10, maybe even a VAResearch box) running Linux
> & Apache would score on SpecWeb.
It wouldn't do very well. Maybe a result of 2000. Linux has scaling
problems on SMP boxes (incoming bandwidth doesn't scale well, outgoing is
fine I believe). Linux' network drivers also won't transition from
interrupt-driven to poll-driven mode under high load, which means cpu time
is wasted in interrupts.
> With the sendfile() stuff, (maybe flow,) some tuning together with the
> Apache Group it probably could be pretty high without httpd in the
> kernel.
No, you really need to get into the kernel to get that high. flow gets
some of the way there, but to go all the way you need a kernel module (you
need flow in the kernel ;).
Dean
Re: record specweb...
Posted by Christof Damian <da...@mediaconsult.com>.
Dean Gaudet wrote:
>
> http://www.sun.com/solaris/xeon/
>
> I'm still having a hard time convincing myself it's not cheating to put
> http into the kernel. ;)
;-)
I wonder how high one of these new four way Xeon boxes (Dell PowerEdge
6300, IBM Netfinity 7000 M10, maybe even a VAResearch box) running Linux
& Apache would score on SpecWeb.
With the sendfile() stuff, (maybe flow,) some tuning together with the
Apache Group it probably could be pretty high without httpd in the
kernel.
Now someone has to convince RedHat to buy a SpecWeb license & Dell/IBM
to provide the hardware.
damian
--
Christof Damian
Technical Director
http://www.mediaconsult.com/ ( btw: mediaconsult is hiring )
RE: record specweb...
Posted by Simon Spero <se...@tipper.oit.unc.edu>.
> > http://www.sun.com/solaris/xeon/
>
> That pretty much destroys their claims about the superiority of
> SPARC hardware. Is a 4-way Xeon really faster than an N-way
> Starfire on SpecWeb?
>
This is almost entirely the result of their software tweaking, though it's
interesting they chose to release results for XEON rather than sparc.
One of the biggest problems with benchmarks like SpecWeb; by the time you've
dumped stuff into the kernel the CPU is basically optional :); there's lots
of ways to optimise for benchmark world (no SSL, no real dynamic pages,
etc). Also, because of the limited number of client streams they don't
reflect how servers would display when handling that number of requests over
a real network (TLB flipping costs and memory cache sizes start to become
more important).
OF course, I remember when I held the blue ribbon for web performance at 130
tps. Of course, that was back in 94 :-)
Simon
RE: record specweb...
Posted by Wesley Felter <we...@cs.utexas.edu>.
> http://www.sun.com/solaris/xeon/
That pretty much destroys their claims about the superiority of SPARC hardware. Is a 4-way Xeon really faster than an N-way Starfire on SpecWeb?
Wesley Felter - wesf@cs.utexas.edu