[lkml]   [1997]   [Jul]   [7]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
Messages in this thread
SubjectRe: InfoWorld web server shootout
I have some problems with their methodolog,y but rather more with their
numbers. Although the annual hardware support costs for a compaq proliant
are fairly easy to calculate they seem to have pulled the $30,000
hardware support number out of the air... Given that you could maintain a
spare box for failover for quite a bit less than that I think that number
is unreasonable.

Although they don't specify the exact hardware for the linux box but they
do specify that it is a dual pentium pro 200 running a uni-processor
kernel (they had trouble with the machine hanging about 1/2 through the
tests with an smp kernel).

There are two quotes that I particularly like:

"These performance numbers consist of URL and image-processing tasks based
on Client/Server Labs' Intramark 1.0 benchmark. The performance numbers do
not include any database server-processing tasks."

"The vendors provided plenty of RAM (256MB)"

The box which we have, a sparc-center 2000e which handles similiar
transient client loads has more like a gig of ram.

They don't seem to have done a whole lot of tuning on the redhat box which
is understandable since they don't seem to be especially technical. I
suspect that this in most cases reflects real life in the sense that most
users aren't going to twiddle with either their kernel or their apache
configuration very much. The rest of the machines come configured more or
less as dedicated web servers where-as the linux box is configured to be a
desktop workstation just as much as it is a server. While I wouldn't
suggest that some random user start in on the 2.1.X kernel series some of
the performance improvements to smp are considerable and should probably
be taken into consideration when testing dual pentium pro machines.

More things to tune would be to adjust a couple of things in the

Timeout -- should be adjusted down from 400
MinSpareServers should be set to something like 20 instead of 5
MaxSpareServers can get really cranked up since with a massive client load
we don't really care if extra servers are hanging around since we'll
need them 100-200
MaxClients should be really cranked (this is why you want lots of ram) say
500 or so would be a good start.

you'd also want to upgrade to 1.2.1 from 1.1.3

At the number of hits they're serving (1.8 million an hour with the aix
box) you'd probably want to tune the networking portion of the kernel
(play with adjusting some buffer sizes) to tune towards high transient
client loads.

As an aside it's hard to tell if the network they constructed is really
condusive to testing any of the machines (they include the cost of a fast
ethernet hub in the cost of the machine) but it seems to me that at a
level of 1920 or so virtual clients building a well switched backbone and
distributing the load between multiple network interfaces (either via
round robin in dns or multiple virtual hosts bound to seperate interfaces)
would allow all of the tested hosts to perform better.


n Mon, 7 Jul 1997, Dave Cinege wrote:

> On Mon, 7 Jul 1997 16:39:44 -0500 (CDT), Michael Brennen wrote:
> >
> >I've cut out most of the article below; the full text is online at the URL
> >listed. Infoworld did a Web server comparison, and to my surprise the
> >Apache / Linux configuration did not do well at all. Considering what it
> >was up against in the Netscape and M$ configurations, something must have
> >been seriously wrong with the Apache / Linux test rig. BTW I don't know
> >how a Linux Hardware Solutions 2000P platform is built.
> Knowing M$ InfoWorld it's a 486-33 with 32mb ram.
> > Microsoft solution
> > Internet Information Server 3.0
> > HP NetServer LX Pro 6/200 SMP
> with 2 3 or 4 PPros??
> > Microsoft solution
> >
> > Total attempted
> > requests 3,888 15,419 23,095 19,043
> Magically M$ wins again. Jeez...why is it they are so good in the mags and
> when I try their products I'm lucky to get them it install?
> ------------------------------------------------------------------
> Elite MicroComputers 908-541-4214

Joel Jaeggli
Academic User Services
"On two occasions I have been asked [by members of Parliament!], `Pray,
Mr. Babbage, if you put into the machine wrong figures, will the right
answers come out?' I am not able rightly to apprehend the kind of
confusion of ideas that could provoke such a question."
-- Charles Babbage

 \ /
  Last update: 2005-03-22 13:39    [W:0.057 / U:11.488 seconds]
©2003-2018 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site