 |
|
|
|
|
|
| |
| |
|
|
|
|
| |
| |
|
|
Given: One Pentium 4 with hyperthreading (2 virtual processors)
Find: Does running two instances of POV-Ray make sense?
Experimental setup:
1 Dell Dimension Precision 360
1 Pentium 4 3.2GHz processor (2 virtual processors)
1.00GB main memory
Windows XP Service Pack 1
Cygwin execution environment
Megapov 1.01 (Megapov 1.0 with a couple of my own patches added)
compiled with GCC 3.3.1-cygwin
POV-Ray 3.5 standard benchmark.pov version 1.02,
and standard benchmark ini file
Procedure:
1. Run the benchmark on a single instance of pov.
A single instance never uses more than 50% cpu time, according to task
manager
2. Run the benchmark simultaneously on two instances of pov.
The two instances are run in two seperate xterm instances, in the same
directory, with the same command line, manually started less than 1
second apart. Neither instance uses more than 50%, and the combined
total is usually 99% or 100%.
3. Note the real time of completion for all runs. POV-Ray reports this
in its statistics page as total time. This is distinct from the CPU time
used as reported by task manager.
Data:
Single instance: Complete in 37min12sec
Double instance: First complete 58min50sec
Second complete 59min17sec
Analysis and discussion:
This is a custom build of megapov, which itself is a distinct version of
POV-ray. The purpose of this experiment is not to compare this version
with any other benchmark. The only variable in this experiment is number
of simultaneously running instances.
To compare the runs most easily, compute pixels per second. On the
single instance, divide the number of pixels rendered by the number of
seconds in the render
384*384= 147456 pixels
37:12= 2232 seconds
147456/2232=66.06pixels/second.
On the double instance, the number of pixels is twice as much, since two
images were rendered, and the render time is the longer of the two runs,
since they were started simultaneously and what matters to me is the
amount of real time between start and end of render.
384*384*2= 294912 pixels
59:17= 3557 seconds
294912/3557=82.91pixels/second
Results:
Running two instances results in an improvement of (82.91/66.06)-1=25.5%
speed improvement. Two identical independent computers can reasonably
expect a 100% improvement.
Conclusion:
Hyperthreading does result in an improvement of rendering speed, but
only 25%. This is small, but speed is speed, and any improvement is
good. It does make sense two run two instances, but do not expect twice
the rendering speed.
Post a reply to this message
|
|
| |
| |
|
|
|
|
| |
| |
|
|
> Hyperthreading does result in an improvement of rendering speed, but
> only 25%. This is small, but speed is speed, and any improvement is
> good. It does make sense two run two instances, but do not expect twice
> the rendering speed.
Since there is a single physical CPU, I'd say the 25% speedup
is actually quite good.
Now the question is: why does a single instance run slower
in the first place ? Or, IOW: where do the two instances really gain
speed (file i/o operations or elsewhere) ?
- NC
Post a reply to this message
|
|
| |
| |
|
|
|
|
| |
| |
|
|
On Thursday 22 July 2004 23:12, Chris Jeppesen wrote :
> Given: One Pentium 4 with hyperthreading (2 virtual processors)
> Find: Does running two instances of POV-Ray make sense?
>
> Experimental setup:
>
> 1 Dell Dimension Precision 360
> 1 Pentium 4 3.2GHz processor (2 virtual processors)
> 1.00GB main memory
> Windows XP Service Pack 1
> Cygwin execution environment
> Megapov 1.01 (Megapov 1.0 with a couple of my own patches added)
> compiled with GCC 3.3.1-cygwin
> POV-Ray 3.5 standard benchmark.pov version 1.02,
> and standard benchmark ini file
>
> Procedure:
>
> 1. Run the benchmark on a single instance of pov.
> A single instance never uses more than 50% cpu time, according to task
> manager
> 2. Run the benchmark simultaneously on two instances of pov.
> The two instances are run in two seperate xterm instances, in the same
> directory, with the same command line, manually started less than 1
> second apart. Neither instance uses more than 50%, and the combined
> total is usually 99% or 100%.
> 3. Note the real time of completion for all runs. POV-Ray reports this
> in its statistics page as total time. This is distinct from the CPU time
> used as reported by task manager.
>
> Data:
>
> Single instance: Complete in 37min12sec
> Double instance: First complete 58min50sec
> Second complete 59min17sec
>
> Analysis and discussion:
>
> This is a custom build of megapov, which itself is a distinct version of
> POV-ray. The purpose of this experiment is not to compare this version
> with any other benchmark. The only variable in this experiment is number
> of simultaneously running instances.
>
> To compare the runs most easily, compute pixels per second. On the
> single instance, divide the number of pixels rendered by the number of
> seconds in the render
>
> 384*384= 147456 pixels
> 37:12= 2232 seconds
>
> 147456/2232=66.06pixels/second.
>
> On the double instance, the number of pixels is twice as much, since two
> images were rendered, and the render time is the longer of the two runs,
> since they were started simultaneously and what matters to me is the
> amount of real time between start and end of render.
>
> 384*384*2= 294912 pixels
> 59:17= 3557 seconds
Add to this the second time
58:50 = 3530
makes a total of 7087 seconds
294912 / 7087 = 41.6 pixels per second
which makes 66.06 / 41.6 = 1.59 speed factor
So a single instance is 1.59 times faster.
And the reason for this is that all non-Povray stuff is done by the other
processor so Povray has one processor all to itself.
>
> 294912/3557=82.91pixels/second
>
> Results:
>
> Running two instances results in an improvement of (82.91/66.06)-1=25.5%
> speed improvement. Two identical independent computers can reasonably
> expect a 100% improvement.
>
> Conclusion:
>
> Hyperthreading does result in an improvement of rendering speed, but
> only 25%. This is small, but speed is speed, and any improvement is
> good. It does make sense two run two instances, but do not expect twice
> the rendering speed.
Your conclusion is a little of because you never considered the total time
needed to do the two simultaneous renders.
I have run simular tests on real dual proc computers and the show the same
results (with not such big differences)
--
Ger
Post a reply to this message
|
|
| |
| |
|
|
|
|
| |
| |
|
|
Ger wrote:
> Add to this the second time
> 58:50 = 3530
> makes a total of 7087 seconds
> 294912 / 7087 = 41.6 pixels per second
> which makes 66.06 / 41.6 = 1.59 speed factor
>
> So a single instance is 1.59 times faster.
>
> And the reason for this is that all non-Povray stuff is done by the other
> processor so Povray has one processor all to itself.
>
I stand by my original conclusion. What matters to me is real live wall
clock time from beginning of render to end. All 58 minutes of one render
were within the 59 minutes of the other.
The single test starts (for instance) at 1:00:00 and finishes at
1:37:12. The double test starts both instances simultaneously at
2:00:00, and ends at 2:59:17 (when the second render finishes). It takes
more time, but not twice as much time. When it is done, it has rendered
twice as many pixels.
I like to do animations, so how I would generalize this is that the
first render is running all the odd frames, and the second is running
all the even frames. While it is true that it will take longer to run
the odd frames simultaneous with the even frames than without, the total
time for all the frames is less when simultaneous than when not.
Post a reply to this message
|
|
| |
| |
|
|
|
|
| |
| |
|
|
Nicolas Calimet wrote:
> Since there is a single physical CPU, I'd say the 25% speedup
> is actually quite good.
> Now the question is: why does a single instance run slower
> in the first place ? Or, IOW: where do the two instances really gain
> speed (file i/o operations or elsewhere) ?
>
> - NC
The benchmark renders directly into the bitbucket, so after the SDL is
parsed and files loaded, there is almost no I/O at all. The system has
1GB of physical memory and the entire OS and all processes in it never
used more than 300MB of it during these tests, so it should not be
swapping. I think it really is hyperthreading helping out at this point.
The theory behind hyperthreading is that decoding instructions is the
current bottleneck in a chip. The decoder and execution units are in a
pipeline, but there is only one decoder and it can only keep some of the
execution units busy. For instance, if the program is running a series
of floating point ops, the integer units are idle.
Hyperthreading addresses this by adding another decoder, visible to the
OS and programs as another set of registers and instruction pointer. It
really does look like another processor. If one decoder is running a
series of floating point instructions, another program running on the
other decoder could run a bunch of integer instructions, keep the
integer units busy, use the chip more efficiently, and crank out twice
as much work in the same amount of time.
This is of course theoretical, since real programs such as Pov use both
kinds of instructions, and keep them fairly evenly mixed. In Pov, all
the ray calculations are done in floating point, but all the address
calculations, array indexing, pointer arithmetic, tree traversal, etc,
use the integer units. If two instances of Pov happened to be running
opposite kinds of instructions, they would run very efficiently. If they
happened to both be crunching integers at the same time, one of the
decoders would not be able to use the integer units since the other
decoder was. It would have to pause until the other was finished. In
this case the decoders run almost in series, and there is little to no
gain in efficiency.
My test suggests that over a long render, Pov is somewhere in between,
but closer to the second case.
It would be interesting to run this same test on a single-processor
non-hyperthreaded system. I suspect that if it really is hyperthreading,
and not working through I/O waits, that causes the speed increase, then
running two instances simultaneously would take more than twice as long
as a single instance.
Post a reply to this message
|
|
| |
| |
|
|
|
|
| |
| |
|
|
I agree with Chris. There's a flaw in your thinking.
--
#macro M(A,N,D,L)plane{-z,-9pigment{mandel L*9translate N color_map{[0rgb x]
[1rgb 9]}scale<D,D*3D>*1e3}rotate y*A*8}#end M(-3<1.206434.28623>70,7)M(
-1<.7438.1795>1,20)M(1<.77595.13699>30,20)M(3<.75923.07145>80,99)// - Warp -
Post a reply to this message
|
|
| |
| |
|
|
|
|
| |
| |
|
|
Chris Jeppesen <pov### [at] kwansystems org> wrote:
> Hyperthreading addresses this by adding another decoder, visible to the
> OS and programs as another set of registers and instruction pointer. It
> really does look like another processor. If one decoder is running a
> series of floating point instructions, another program running on the
> other decoder could run a bunch of integer instructions, keep the
> integer units busy, use the chip more efficiently, and crank out twice
> as much work in the same amount of time.
So in theory if a program needs to perform floating point calculations
and (independent) integer calculations, it should run those calculations
in separate threads for maximum speed advantage in a P4?
--
plane{-x+y,-1pigment{bozo color_map{[0rgb x][1rgb x+y]}turbulence 1}}
sphere{0,2pigment{rgbt 1}interior{media{emission 1density{spherical
density_map{[0rgb 0][.5rgb<1,.5>][1rgb 1]}turbulence.9}}}scale
<1,1,3>hollow}text{ttf"timrom""Warp".1,0translate<-1,-.1,2>}// - Warp -
Post a reply to this message
|
|
| |
| |
|
|
|
|
| |
| |
|
|
That would be?
--
Ger
Post a reply to this message
|
|
| |
| |
|
|
|
|
| |
| |
|
|
Ger <ger### [at] hotmailcom> wrote:
> That would be?
You are basically saying that rendering two benchmarks using two
threads should be faster than rendering one benchmark using one thread.
This is, of course, just flawed thinking. There's no way of getting
two benchmarks rendered faster even when using two distinct processors
than one benchmark in one processor.
What he is saying is that rendering two benchmars at the same time
gets rendered 25% faster than rendering first one benchmark and then
the other. The end result of both test is identical, but the first
test produced it 25% faster.
--
plane{-x+y,-1pigment{bozo color_map{[0rgb x][1rgb x+y]}turbulence 1}}
sphere{0,2pigment{rgbt 1}interior{media{emission 1density{spherical
density_map{[0rgb 0][.5rgb<1,.5>][1rgb 1]}turbulence.9}}}scale
<1,1,3>hollow}text{ttf"timrom""Warp".1,0translate<-1,-.1,2>}// - Warp -
Post a reply to this message
|
|
| |
| |
|
|
|
|
| |
| |
|
|
Chris Jeppesen wrote:
> Given: One Pentium 4 with hyperthreading (2 virtual processors)
> Find: Does running two instances of POV-Ray make sense?
Your testing is very interesting.
How about running the exact same tests with hyperthreading turned OFF?
So you'll have 4 results on the exact same hardware
single - with hyperthreading
double - with hyperthreading
single - no hyperthreading
double - no hyperthreading
It would be interesting to see the difference between the hyperthreading
and non-hyperthreading results.
This might help show what you have found through your testing.
Tom
Post a reply to this message
|
|
| |
| |
|
|
|
|
| |
| |
|
|
On Friday 23 July 2004 14:49, Warp wrote :
> Ger <ger### [at] hotmailcom> wrote:
>> That would be?
>
> You are basically saying that rendering two benchmarks using two
> threads should be faster than rendering one benchmark using one thread.
Either I said it wrong or you understood wrong.
Rendering 1 image on a computer with more then one processor will always be
faster then rendering the same on a single(non-hyper) processor.
What I was (tried to) pointing out that the total time spend on a single
pixel in the single instance is less then in the dual instance because of
the non_pov stuff going on.
>
> This is, of course, just flawed thinking. There's no way of getting
> two benchmarks rendered faster even when using two distinct processors
> than one benchmark in one processor.
>
I never said it would. (and if then I definitely need to rephrase that)
> What he is saying is that rendering two benchmars at the same time
> gets rendered 25% faster than rendering first one benchmark and then
> the other. The end result of both test is identical, but the first
> test produced it 25% faster.
>
Sure, I understood that. All I pointed out is that the cpu-time/pixel in the
single instance is less. Otoh, making a statement like this on a
hyperthreading processor is dodgy at best because a hyperthreading
processor is not nearly the same as two distinct processors.
--
Ger
Post a reply to this message
|
|
| |
| |
|
|
|
|
| |
| |
|
|
Ger <ger### [at] hotmailcom> wrote:
> Sure, I understood that. All I pointed out is that the cpu-time/pixel in the
> single instance is less.
And that's where you are wrong. You need to count the total number
of pixels calculated and divide it by the time it took to calculate
all those pixels. In this case it took 56 (or whaver it was) minutes
to render 2xbenchmark pixels.
--
#macro M(A,N,D,L)plane{-z,-9pigment{mandel L*9translate N color_map{[0rgb x]
[1rgb 9]}scale<D,D*3D>*1e3}rotate y*A*8}#end M(-3<1.206434.28623>70,7)M(
-1<.7438.1795>1,20)M(1<.77595.13699>30,20)M(3<.75923.07145>80,99)// - Warp -
Post a reply to this message
|
|
| |
| |
|
|
|
|
| |
| |
|
|
On Friday 23 July 2004 19:40, Warp wrote :
> Ger <ger### [at] hotmailcom> wrote:
>> Sure, I understood that. All I pointed out is that the cpu-time/pixel in
>> the single instance is less.
>
> And that's where you are wrong. You need to count the total number
> of pixels calculated and divide it by the time it took to calculate
> all those pixels. In this case it took 56 (or whaver it was) minutes
> to render 2xbenchmark pixels.
>
Difference is that you are talking about "total computer" time, I was
talking about "per processor" time.
--
Ger
Post a reply to this message
|
|
| |
| |
|
|
|
|
| |
| |
|
|
> Difference is that you are talking about "total computer" time, I was
> talking about "per processor" time.
... which is the same since there's a single CPU in the computer :-)
- NC
Post a reply to this message
|
|
| |
| |
|
|
|
|
| |
| |
|
|
On Friday 23 July 2004 21:02, Nicolas Calimet wrote :
>> Difference is that you are talking about "total computer" time, I was
>> talking about "per processor" time.
>
> ... which is the same since there's a single CPU in the computer :-)
>
> - NC
Not entirely true. It's not the same as a very fast Pentium II. Nor is it a
very fast dual Pentium II. But it's enough of a dual proc to make a
difference when running two parallel renders.
--
Ger
Post a reply to this message
|
|
| |
| |
|
|
|
|
| |
| |
|
|
Ger <ger### [at] hotmailcom> wrote:
> Difference is that you are talking about "total computer" time, I was
> talking about "per processor" time.
Who cares about the "per processor" time?
People want their renders to finish as fast as possible.
--
#macro N(D)#if(D>99)cylinder{M()#local D=div(D,104);M().5,2pigment{rgb M()}}
N(D)#end#end#macro M()<mod(D,13)-6mod(div(D,13)8)-3,10>#end blob{
N(11117333955)N(4254934330)N(3900569407)N(7382340)N(3358)N(970)}// - Warp -
Post a reply to this message
|
|
| |
| |
|
|
|
|
| |
| |
|
|
On Saturday 24 July 2004 10:00, Warp wrote :
> Ger <ger### [at] hotmailcom> wrote:
>> Difference is that you are talking about "total computer" time, I was
>> talking about "per processor" time.
>
> Who cares about the "per processor" time?
> People want their renders to finish as fast as possible.
>
Okay, fair enough :)
--
Ger
Post a reply to this message
|
|
| |
| |
|
|
|
|
| |
| |
|
|
Tom Austin nous apporta ses lumieres ainsi en ce 23/07/2004 09:03... :
> Chris Jeppesen wrote:
>
>> Given: One Pentium 4 with hyperthreading (2 virtual processors)
>> Find: Does running two instances of POV-Ray make sense?
>
>
> Your testing is very interesting.
>
> How about running the exact same tests with hyperthreading turned OFF?
>
> So you'll have 4 results on the exact same hardware
> single - with hyperthreading
> double - with hyperthreading
> single - no hyperthreading
> double - no hyperthreading
>
> It would be interesting to see the difference between the
> hyperthreading and non-hyperthreading results.
>
> This might help show what you have found through your testing.
>
>
>
> Tom
>
I saw some P4 benchmarks using the same CPU with and without
hyperthreading enabled. Most where SLOWER when using hyperthreading.
Sory, I don't remember where I saw those.
Alain
Post a reply to this message
|
|
| |
| |
|
|
|
|
| |
|
|
