Hello group !

Lastly I got myself access to a HP DL360 g9 server and thought i'd give it a try
with povray.

It has : two E5-2430 v2 CPUs @ 2.5 Ghz, each 6 core, making 12 physical cores,
24 logical with HT enabled.

The previous system i used was a BL460c G1 blade with two E5450 @ 3 Ghz CPUs,
each 4 core, making 8 physical cores, no HT.


The fact is that some custom scenes are rendering far slower on the new
platform.


There are certainly a lot of architectural differences between these two kinds
of CPUs, since there is more than 6 years gone beetween the two models.

I may think that the difference in performance may be linked to the small L2
cache of the newer CPU.

You can see all the CPU specs differences in :

http://cpuboss.com/cpus/Intel-Xeon-E5450-vs-Intel-Xeon-E5-2430-v2

When comparing using biscuit.pov it is faster on the new E5-2430-v2 CPUs

But on my custom scene that uses quite a lot of transparencies it is far slower
(2 to 3 times slower) on the new CPUs

Any idea on what may cause such dramatic performance swings ?
Does a custom build could make some performance improvements ?

I am quite new in povray benchmarking so i'd liste to your guidance.
May post detailed benchmarks with povray if needed.

Post a reply to this message

Am 25.04.2015 um 13:15 schrieb rodv92:

> Lastly I got myself access to a HP DL360 g9 server and thought i'd give it a try
> with povray.
>
> It has : two E5-2430 v2 CPUs @ 2.5 Ghz, each 6 core, making 12 physical cores,
> 24 logical with HT enabled.
>
> The previous system i used was a BL460c G1 blade with two E5450 @ 3 Ghz CPUs,
> each 4 core, making 8 physical cores, no HT.
>
>
> The fact is that some custom scenes are rendering far slower on the new
> platform.
>
>
> There are certainly a lot of architectural differences between these two kinds
> of CPUs, since there is more than 6 years gone beetween the two models.
>
> I may think that the difference in performance may be linked to the small L2
> cache of the newer CPU.

As far as the raw number crunching power is concerned, the new computer 
should perform somewhat better: The old one would be equivalent to a 
single 24GHz core, while the new one would be equivalent to a single 
40GHz core, plus a guesstimated 25% bonus from hyperthreading, i.e. a 
50GHz core.

However, there are two caveats to consider:

(1) With the computing power spread across more cores, the new computer 
introduces more synchronization overhead. While most of the computations 
in POV-Ray scale quite well with number of cores, Radiosity in 
particular doesn't play that nice, so in radiosity scenes the higher 
number of cores equals a noticeable loss of effective computing power.

(2) POV-Ray doesn't take the numbers to crunch out of thin air; there's 
the internal representation of the scene as a data source, an image 
buffer as a data sink, and various data buffers - again most notably the 
radiosity buffer - as both sinks and sources. All this data is accessed 
on a regular basis during rendering, and the more of it fits into faster 
caches, the better the CPU can bring its raw power to bear.


> Any idea on what may cause such dramatic performance swings ?

I'd go with your guess that the significantly smaller L2 cache of the 
new computer is the most important factor. If the data and buffers for a 
scene fit entirely inside the new computer's L2 cache, it will certainly 
outperform the old one, but with more memory-hungry scenes it will spend 
much of its time waiting for data from the main memory rather than 
actually computing anything.

> Does a custom build could make some performance improvements ?

That's rather unlikely. Not with the problematic scenes, that is. A 
custom build that utilizes the new CPU's instruction set extensions 
might improve the speed at which the CPU can crunch the data, but it 
won't be able to significantly improve the speed at which the data can 
be fetched from memory in the first place.

Post a reply to this message