On Tue, Jul 25, 2000 at 01:23:18PM -0600, Bruce Guenter wrote:
> I have written a benchmark that iterates over message sizes from 1000 to
> 64000 bytes, and from 1 to 16 recipients, and times how long it takes to
> send the same message to all the recipients using qmail-remote. It
> calls qmail-remote once with all the recipients (multi-RCPT), and once
> for each recipient (multi-connection). I only have preliminary results
> so far, and I plan to run a more complete set of tests tonight after I
> leave work. I'll post my full results and scripts once I've completed
> the tests.
As promised, I've posted the results of the benchmark testing at
http://em.ca/~bruceg/bench-qmail-remote/
The receiving server is my PC, which has a DSL connection running at
about 1.5Mb downlink bandwidth (the part that was actually used) running
qmail, of course. The "-cable-" results were sent from a cable modem
which has approximately 384Kb uplink bandwidth. the "-2Mb-" results
were sent from a partial DS3 with 2Mb of bandwidth. The receiver had
its concurrency set to 128.
20 runs were done of each test, 10 with one connection with multiple
recipients, and 10 with multiple connections with one recipient. The
min and max columns give the fastest and shortest run times
respectively; mean is (T1*T2*T3...*T10)**(1/10); avg is
(T1+T2+T3+...+T10)/10. The mean is less biased by unrepresentative
results, and so is a better measure of the common case.
Conclusions are somewhat tricky. Using mutiple RCPTs tends to be more
predictable (less of a spread between min and max), but using multiple
connections has the best optimistic behaviour (min is lower than
multi-RCPT's min). With small messages (4KB and less), multi-connection
is always a win. On our mail proxy, the median message size is 3KB,
just for comparison. On the well-connected sender, using multi-RCPTs
was never a significant win, which proves DJB's hypothesis about its use
for well-connected hosts. Once bandwidth limits become an issue (poorly
connected server, large messages), multi-RCPTs win because the latency
involved in sending one more RCPT becomes less than the additional time
required to send another concurrent copy.
This says nothing about bandwidth efficiency, only time efficiency.
Obviously, using multi-RCPTs is always a bandwidth win (unless your
recipient is larger than your message, highly unlikely).
Feedback would be appreciated. Oh, and please don't consider the test
addresses I used in the scripts as wide open for mailbombing.
--
Bruce Guenter <[EMAIL PROTECTED]> http://em.ca/~bruceg/
PGP signature
