From: Greg Ercolano <erco@(email surpressed)>
Subject: Re: HyperThreading
   Date: Thu, 24 Mar 2011 17:22:08 -0400
Msg# 2062
View Complete Thread (2 articles) | All Threads
Last Next
Daniel Browne wrote:
> We recently added some new render nodes to our farm which came with HyperThreading
> disabled by default in their BIOS. I was curious as to everyone's experiences
> with and opinions on the advantages and drawbacks of having HyperThreading
> enabled? I know that before the Nehalem series CPUs it was not
> considered to provide a practical benefit for compute intensive tasks like
> rendering.

	Others feel free to chime in.. here's a thread on hyperthreading
	from the "studio sysadmins" group <http://studiosysadmins.com/>:
	http://www.studiosysadmins.com/search/?search_term=hyperthreading?

	You should join; it's a good group with a focus on VFX sysadmining..
	very tech issues covered.

	I've excerpted these replies (in case the link goes stale):

* * *
[Stephen,primefocus/uk]
We did some PRMan tests ages ago and hyperthreading off was the better
option.  Has anyone done any recent testing on whether it's a good thing
or not with current procs?
* * *
[Dave,themill]
We're pure mental ray over here and we turn hyperthreading off off off
all the time all the time all the time. draws less current (on most of
our machines) and renders faster. wins all around
* * *
[Dan,themill]
In all my testing with the majority of kit - I'd say hyperthreading off
is the best go. Naturally this will depend product to product and what
you're utilising inside of each renderer.

MR and PRMan have def. been better with it off in my experience,
but we're majority 54xx here - not a worry! :)
* * *
[Vince/bluesky]
Test your specific frames and code with it on and off. You will find
some code will go faster, some will go slower. Also, if it's HT from
prior generations, even Intel said it was broken. New procs are much
better. If your code doesn't like HT, it *will* slow down your renders
(just like bad multi-threading code will, on a per-core basis).
* * *
[Jason,?]
Windows 7 has some added abilities to it which allows Hyperthreading
to work better. Basically it can give priority to a physical processor
over a virtual one with improved logic from XP.
http://www.tomshardware.com/news/windows-hyperthreading-intel-nehalem-atom,7831.html
and 30 seconds into this video there is a small bit about it as well:
http://software.intel.com/en-us/videos/windows7-intel-microsoft-consumer-business-benefits/?wapkw=(hyperthreading+windows+7)
* * *
[Jeff]
Agreed, hyperthreading off is the way to go. Also found that our
version of Shave and Haircut wouldn't render properly w/
hyperthreading enabled for MR.
* * *
[Douglas]
We run mostly V-Ray and Maxwell on our farm, with HT definitely ON.
Maxwell in particular gets a huge bump in performance.
* * *
[Christopher]
Vue renderer too. Not sure about the new version but we were getting
up to 40% gains HT on in V8. I'd assume this relates to how
efficiently threaded the code is as our more "mature" renderers see
little or no gain.

It's off anyway as I specified processors which, with HT on, exceed
the thermal capacity of our node hardware. Vendor didn't catch the
mistake either. Live and learn.

Maxwell for animation? Or are you splitting frames? How's that working
out? Love the results but couldn't imagine it's use, at least on our
teevee schedule...

[Douglas responds: "A mix of product design and animation for broadcast.
More VFX-y stuff is usually V-Ray or MR."]

* * *
[David,pixar]
I think it is important to specify which processor family you are
using, and probably which renderer versions too.  Certainly prman
didn't gain much from the "original flavor" of hyperthreading
on the P4/Xeon family.

On the new "Nehalem" architecture (i3,i5,i7), we are seeing a
measurable benefit, sometimes significant, from the hardware
"threading" mode (Intel no longer terms it "hyperthreading").
It does appear to be most effective when the app is running in
a many-threaded mode.  And, not surprisingly, prman-16.0 (now
in beta) performs better than prior versions.  The feature does
add some heat and draws some additional power, I don't know
enough about it to characterize the actual overhead.

We do know of studios that actively disable the hardware threading
because they are running most of their apps in single-threaded mode.
The hardware threading mode may or may not be easy to disable when
doing your own benchmark testing; I'm told that it is particularly
obscure to disable on OSX.

-- 
Greg Ercolano, erco@(email surpressed)
Seriss Corporation
Rush Render Queue, http://seriss.com/rush/
Tel: (Tel# suppressed)ext.23
Fax: (Tel# suppressed)
Cel: (Tel# suppressed)

Last Next