At 12:16 -0400 2003/10/05, Richard Welty wrote: >On Sun, 5 Oct 2003 16:45:32 +0200 Giuliano Gavazzi
><eximlists@???> wrote:
>> even without SQL, how would you propose a "real life" test to be
>> done? Where will you get all the hosts to receive from and delivery
>> to, with all the DNS lookups and network delays? Perhaps we should
>> have a sort of contest, where list members will give a list of
>> addresses available to send to and will also be given a list of
>> addresses on the test systems to send to.
>
>that would be too many uncontrolled variables, i don't know that the
>numbers generated would have any real value.
>
>something organized and structured, a carefully designed "torture test"
>sequence, which only varies one variable at a time, would seem to be in
>order.
>
>doing it properly would take a _lot_ of work.
I agree with what you say, but still I think the "real life"
environment is quite important. Unless the lifetime of processes is
not important, just banging a machine with messages from a couple of
hosts on a fast LAN does not seem satisfactory. Fortunately real life
conditions can be simulated on a LAN (for instance ipfw pipes for b/w
limitation and delays).
I would then keep this "heat bath" constant (the network
configuration and the "client" machines, meaning dns server and mail
peers) while changing conditions on the test machine.
A quick rule for the network config would be, keep the ping times at
sensible values (that is 80/100ms rather than 3ms) and allow the
bandwidth to each simulated peer to be equal to 2MB (an E1) divided
by the average number of simultaneous connections.
The design would still be quite a feat, at least a good exercise in networking!
