On Thu, 11 Apr 2002, Eduard Tieseler wrote:
> Can anyone tell me the amount of traffic they are passing with Pen? I
> installed and got it working in about a minute on a test network but need it
> to work on a mission critical system. It works fine, but I am expecting
> over 100,000 hits a day. Has any one abused Pen with allot of traffic?
For a client of ours, I did some testing using this setup:
.------------. .--------. .--------.
| APACHE+PEN | | APACHE | | APACHE |
`------------´ `--------´ `--------´
| | |
--------------------------------
| |
.-----. .-------.
| NFS | | MYSQL |
`-----´ `-------´
As you see, pen is running on one of the web servers. All the software
and all the web pages are on the nfs server. The database wasn't used
in this test.
All servers have single 1 GHz Celerons, 512 MB RAM and Fast Ethernet
nics (Intel 82559 on the motherboard). They are connected to a
Cisco 3548 switch.
Slackware Linux 8.0 on everything, using the standard 2.4.5 kernel.
I ran pen in blocking mode, i.e. using the -n option. My experience is
that pen runs much faster and lighter in that mode, and it will be the
default in the next release.
Used Jef Poskanzer's http_load on the nfs and mysql servers like this:
http_load -parallel 100 -seconds 60 urls
Both clients got about 500 fetches per second; the nfs slightly less,
the mysql server slightly more. That's reasonable since the nfs server
was serving files to the web servers during the test.
So in this test I got a total of 1000 fetches per second. Nothing was
maxed out. Pen used about 9% of the CPU and its usual 500 K of RAM.
A few comments:
Nothing was tuned in this setup. I like to do initial benchmarks
on untuned systems just to get a baseline. Otherwise there is no way
to know if there is any benefit from the tuning.
I run pen on one of the web servers. I kind of recommend that for
anything but really heavy load, and then I would actually go for
a hardware solution instead. 100000 hits per day doesn't qualify as heavy
load. ;-)
Since there were only two clients, only two of the web servers were
used. A better test would have hundreds of relatively slow clients.
Ulric
This archive was generated by hypermail 2.1.2 : Fri Apr 12 2002 - 10:19:49 CEST