Skip site navigation (1)Skip section navigation (2)
Date:      Mon, 5 Oct 1998 17:45:58 -0700 (PDT)
From:      Marc Slemko <marcs@znep.com>
To:        Eivind Eklund <eivind@yes.no>
Cc:        Reginald Perry <perry@zso.dec.com>, freebsd-chat@FreeBSD.ORG
Subject:   Re: PC Magazine 10/20/1998 Article about FreeBSD
Message-ID:  <Pine.GSO.4.02A.9810051727420.22125-100000@redfish>
In-Reply-To: <19981005222711.23452@follo.net>

next in thread | previous in thread | raw e-mail | index | archive | help
On Mon, 5 Oct 1998, Eivind Eklund wrote:

> On Mon, Oct 05, 1998 at 12:12:50PM -0700, Reginald Perry wrote:
> > Hi there,
> > 	There is an article in the Net Tools, From The Bench section of PC Magazine
> > talking about FreeBSD 2.2.7. Looks pretty factual, but there was one
> > confusing statement. They initially configured both machines with 128MB of
> > RAM. They then increased the RAM and noted that as you do this NT surpasses
> > FreeBSD in their performance measure. They state that this is because of a
> > cache limitation in Apache and FreeBSD. Is this true? Could someone describe
> > this in more detail if so?
> 
> FreeBSD is tuned to have max performance when it get under load - ie,
> when it actually is doing something.  There should not be any
> limitations to the use of cache - FreeBSD basically regard
> _everything_ as cache.  Your entire RAM is just a cache for the disk.
> 
> I'd guess the benchmark interpretation comes from the reviewer doing a
> wild guess on why FreeBSD was slower.

Partly.  

In my experience, zdnet has had this odd assumption for the past year or
so that web servers must have a large in memory cache or else they can't
be fast.  From what I have been told, both by people there and from other
vendors, is that when they went from webbench 1.1(?) to 2.0, the size of
the test set increased a lot.  One of Netscape's servers had been using a
fixed size cache containing a copy of the data to get good benchmark
numbers; with the smaller working set of documents in 1.1, this worked
well.  Then they went to 2.0, their cache was too small so it resulted in
massive thrashing and very slow numbers.  When Netscape ran back to the
office and compiled a new DLL with a bigger cache size that could hold the
document tree in memory and threw it in (saying something about fixing the
cache), performance shot up.  Therefore, certain people at zdnet decided
that you always need to have a seperate in memory cache.  Michael Johnson
(Red Hat) and I tried to get them to understand that the OS is the thing
that should be doing the caching, and while some people appeared to get
the hint, others probably still don't have the hint.  The environment also
sucks clueons away very quickly.

Who are the authors on the review?  

Note that in the typical zd setup, the reviewers are in New York but the
benchmarking is done somewhere else (often the Bay area) and the zdbench
people that write their testing software is in yet another location.  So
you have these massive three way (or four, with a vendor) conference calls
on really pathetic phones.

Now, as to the reason why it dropped off, there are several possibilities.

First, MaxClients was set too low.  Unlikely, since someone posted that
they only went up to 60 clients, which is typical (I won't get started on
exactly what sort of environment such benchmarks simulate and how I have
seldom seen anything even close to that in real life) of zdnet.

Second, they had low process limits.  If they installed Apache themself or
if the port is bogus (dont' know, haven't looked), things like low
processes per user limits could certainly hurt this.

Now, there _is_ some validity to the suggestion that IIS does better
caching.  With Apache, you have to get the metadata about each file again
for each hit, you have to open and close the file for each hit, etc.  IIS
essentially keeps a cache of open descriptors and has a few optimizations
to let NT pump the data from the descriptor to the network very quickly. I
would like to see them do a comparison for dynamic in-process content
using an Apache module vs. ISAPI.  (They don't have an Apache module in
the current webbench distribution so won't run tests using Apache modules,
and when I keep asking over and over for them to include one, even
offering a pre-written one that does just what they need, I get no
response).  With IIS, using a module for dynamic content (very trivial
dynamic content; essentially a printenv) results in a fair performance
drop.  With Apache, it could well result in a performance improvement.

I am suprised that they didn't contact anyone at FreeBSD; it would be
useful for someone there to get in touch with them and gently remind them,
for the future, that people are available to help.  Vendor help in
configuration is the norm.


To Unsubscribe: send mail to majordomo@FreeBSD.org
with "unsubscribe freebsd-chat" in the body of the message



Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?Pine.GSO.4.02A.9810051727420.22125-100000>