httpd-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Marc Slemko <>
Subject Re: another performance paper
Date Mon, 24 Feb 1997 02:54:13 GMT
On Sun, 23 Feb 1997, Brian Behlendorf wrote:

> Actually the paper is linked off of this, gzip'd postscript.

Well... a few interesting bits, nothing really new and completely
worthless as a scientific paper because they give hardly any useful
information about their tests.  Not only are they vague on their test
methodology, but they don't even say what versions of the various servers
they used nor anything about how they are configured.  I can make Apache
outperform any server and make any server outperform Apache.  They don't
even provide numeric results, just hard to read 3d graphs.  Their tests
don't even begin to be reproducable. 

I find their numbers of 50% of time spent forking for Apache to be crazy
unless they have it configured incorrectly for their load.  Certainly if
you set StartServers to 1 and then try to run a benchmark from a freshly
started server you will have a huge slowdown at the start, but that
doesn't say much.

I also question their use of latency measurements; on their test, it is
possibly for a server to get better results by starting off sending as
soon as it gets the request; heck, it can do so before it even reads the
request but that doesn't mean anything performance wise.  There is some
validity in latency measurements, but for current clients (and even
high-speed networks) more analysis is needed to make such numbers have any

They are also weak on implmentation details on their server, making me
wonder if they cut some corners to make things perform better but which
leaves out functionality which is critical in real life.

I'm not saying that Apache is the best performer (although I would have to
question some of their results, depending on the version they used) but I
have trouble taking this paper seriously; what they are saying may be
wonderful research, but they provide nothing to back it up.

If you took out the worthless (ie. unreproducable and unsupported) 
numbers, this would look to me more like an abstract than a paper, but
that's not how they are presenting it.

> BTW, anyone got a good postscript viewer for Win95/NT?


Sorry, all I do when I see a NT box is install FreeBSD... so FreeBSD is a
good postscript viewer.

View raw message