Results 1 to 5 of 5

Thread: Linux, FreeBSD, Solaris i7-3960X Scaling Performance

  1. #1
    Join Date
    Jan 2007
    Posts
    14,361

    Default Linux, FreeBSD, Solaris i7-3960X Scaling Performance

    Phoronix: Linux, FreeBSD, Solaris i7-3960X Scaling Performance

    Using the new Intel Core i7 3960X Extreme Edition Sandy Bridge processor, Scientific Linux 6.1, Debian GNU/Linux, Debian GNU/kFreeBSD, and Solaris 11 11/11 were benchmarked when having a different number of CPU cores enabled to see how well each operating system scales up to six cores plus Hyper Threading.

    http://www.phoronix.com/vr.php?view=16848

  2. #2

    Default

    Michael,

    "Time to compile" - "more is better" - really? WTF?

  3. #3

    Default

    Quote Originally Posted by birdie View Post
    Michael,

    "Time to compile" - "more is better" - really? WTF?
    All results are normalized as said in the article so on those graphs, yes, better is faster.

  4. #4
    Join Date
    Jan 2009
    Posts
    1,305

    Default Check out the Phong rendering graph.

    Is solaris scaling TOO well here?
    At the 2 core mark it looks like it is above 2 times better.
    I suppose the code could be written in such a way that it has too much overhead for small number of processors, but does anyone have any experience writing superscalable code that can say for sure?

  5. #5
    Join Date
    Jan 2010
    Posts
    3

    Default Re: Check out the Phong rendering graph

    (Disclaimer: I am the author of the renderer)

    Indeed, I noticed that in the graph, too - but there's a simple explanation.
    The renderer is using simple OpenMP loops, of the form:

    #ifdef USE_OPENMP
    #pragma omp parallel for
    #endif
    // Plot all the vertices of object i on the canvas
    for(int j=0; j<(int)_vertices.size(); j++) {

    ...so no, in theory it is impossible to see a speedup of more than 2x when using 2 threads instead of one - or a speedup of more than 4x when using 4 threads.

    Unless... the single-core case is handled BADLY by Solaris.

    Michael plotted the relative speedup when going from one core to 2, 4, 6, etc.
    In this case, the single-core running speed of Solaris ( as reported in the full benchmark results) is 21.06 frames per second, when Linux scored around 28 frames per second - in both cases, using a single core!

    So, in effect, what we are seeing here is that Solaris is "punishing" single-core executions - i.e. the OpenMP library in Solaris handles single-core machines very badly.

    P.S. Perhaps this affected other benchmarks too - it's easy for Solaris to appear to have better scalability, when the single-core performance is much worse than the other contenders (Linux and BSD in this case). For example, look at the speed of C-Ray when run with a single core under Solaris vs the other two (Linux/BSD) - it is about half as much (250 seconds vs 500 seconds). Solaris "punishes" single-core executions so much, that it appears to scale much better as more cores are introduced...

    Thanassis.
    Last edited by ttsiodras; 12-20-2011 at 11:28 AM. Reason: In hindsight, other benchmarks might be affected from this, too...

Posting Permissions

  • You may not post new threads
  • You may not post replies
  • You may not post attachments
  • You may not edit your posts
  •