OK, I feel silly having to ask this, but I'm paranoid enough that I will
anyway.
The company I work for is about to plop down a very large amount of money
on bigger, better machines to handle the work. We're currently planning on
an 8-way machine running postgres for the database work, and a group
of dual CPU web servers running the CGI applications.
Before we get up to our necks in hardware purchases, I'd like to have
some peace of mind that the performance isn't going to decrease horribly
because of the bandwidth limitations in the ethernet connections.
To describe the setup, the database server will be connected via a
gigabit fiber cable to a switch, with 100-megabit connections to each of the
application servers. So, until we get a larger number of app servers, the
limiting factor will be the 100 megabit connections to each individual
server.
I'd like to feel sure that transmitting the result sets over the ethernet
isn't going to cause a significant decrease in response time. After working
through the numbers, I don't think that it will - but I'd love to hear from
those who have done similar setups.
We enabled some logging in the database library that we use, and here is
some information on our result sets:
average result set size: 1.4 kilobytes
maximum result set size: 633 kilobytes
Of those results, here's a breakdown:
Size Frequency
-------------------
0-8K 96%
8K-17K 2%
17K - 24K 0.02%
and it goes down from there.
So, it seems that it would take a very large number of transactions per
second to saturate either the 100 mbit link to the servers, or the gigabit
pipe to the server.
Am I correct, or should I reevaluate? Any comments are welcome.
steve