This came up several times in the past, and I finally got around to
testing my dual Xeon server with hyperthreading turned off. I had read
in some places (like Tom's Hardware I believe), that certain
multi-threaded server applications did not perform as well with
hyperthreading as without hyperthreading. You will see from the
attached graph that this test shows that turning off hyperthreading is
not the way to go with at least the 2.x version of Declude.
Just to explain the graph, at about 7:30 a.m. I turned off
hyperthreading and rebooted the computer. There was an immediate large
spike in processing. I then later rebooted the server because this
spike was so high that I thought that maybe there was something wrong
with the state of the server, but that wasn't the case, it just couldn't
handle the traffic very well even though it was normal in every other
sense. Normally my server cruises along with peak average hourly
utilization of about 40% to 45% on weekdays, but today it reached a peak
hourly utilization of 80%. Then to make matters even worse, IMail SMTP
crashed at around 7 p.m. (IMail 8.15 HF2). I would imagine that this
had to do with mystery heap instability, though I have Declude set to
just 20 threads which should have been fine.
I was actually expecting little or no change in my results from this
test, and the only explanation that I can think of would be that because
so many single-threaded applications are being used, that managing these
threads represented a large amount of overhead to the server...as large
as the E-mail itself. By having 4 CPU's seen by the system instead of
just 2, it seems to leap past this bottleneck. This also leads me to
believe that as CPU utilization rises, efficiency goes down. This seems
to be my experience at least.
Note that Declude 3.x and IMail 8.2+ might show very different results,
though I expect that they will be similar since much of the processing
goes to the virus scanners and external tests that are plugged into
Declude, and not Declude or IMail themselves. I'll probably test that
out when I make the leap.
Matt
