Find the answer to your Linux question:
Results 1 to 10 of 10
Hi, i search for the memery mapping for ccNUMA, i.e. which cores take which piece of main memery. Can anybody help me? Thanks...
Enjoy an ad free experience by logging in. Not a member yet? Register.
  1. #1
    Just Joined!
    Join Date
    Mar 2011
    Posts
    9

    ccNUMA


    Hi,
    i search for the memery mapping for ccNUMA, i.e. which cores take which piece of main memery.
    Can anybody help me?

    Thanks

  2. #2
    Linux Guru Rubberman's Avatar
    Join Date
    Apr 2009
    Location
    I can be found either 40 miles west of Chicago, in Chicago, or in a galaxy far, far away.
    Posts
    11,753
    Most PC type systems these days are not NUMA architectures (excepting some AMD Opteron systems), so all cores share all memory. What kind of hardware are you running? From the Wikipedia article on this subject:

    Current ccNUMA systems are multiprocessor systems based on the AMD Opteron, which can be implemented without external logic, and Intel Itanium, which requires the chipset to support NUMA. Examples of ccNUMA enabled chipsets are the SGI Shub (Super hub), the Intel E8870, the HP sx2000 (used in the Integrity and Superdome servers), and those found in recent NEC Itanium-based systems. Earlier ccNUMA systems such as those from Silicon Graphics were based on MIPS processors and the DEC Alpha 21364 (EV7) processor.
    Sometimes, real fast is almost as good as real time.
    Just remember, Semper Gumbi - always be flexible!

  3. #3
    Just Joined!
    Join Date
    Mar 2011
    Posts
    9

    Nehalem

    I work on workstation.
    Opteron has ccNuma-Architecture as AMD said. But i found it isn't.
    Has the linux-kernel such data about the physical memory mapping?

  4. $spacer_open
    $spacer_close
  5. #4
    Linux Guru Rubberman's Avatar
    Join Date
    Apr 2009
    Location
    I can be found either 40 miles west of Chicago, in Chicago, or in a galaxy far, far away.
    Posts
    11,753
    Quote Originally Posted by cswb View Post
    I work on workstation.
    Opteron has ccNuma-Architecture as AMD said. But i found it isn't.
    Has the linux-kernel such data about the physical memory mapping?
    The Opteron chip is capable of NUMA/ccNUMA without additional support chips, but it isn't necessarily used that way. You need an appropriate operating system, or system extensions, to do that. Since ccNUMA is usually used for distributed cluster systems, you may have that ability if you were to run Linux in a cluster. However, on a single board, even with multiple CPUs and cores, it probably isn't going to be running in that mode. The 'cc' part is Cache Coherence, which means when you have physically separated CPUs as in a cluster, it will keep their cache in sync, albeit with some performance degradation, which is why clusters run in this mode will have very high speed interconnects (10gbps would be good).
    Sometimes, real fast is almost as good as real time.
    Just remember, Semper Gumbi - always be flexible!

  6. #5
    Linux Guru Rubberman's Avatar
    Join Date
    Apr 2009
    Location
    I can be found either 40 miles west of Chicago, in Chicago, or in a galaxy far, far away.
    Posts
    11,753
    So, why do you want to do this? Are you building an HPC (high performance computing) cluster, personal supercomputer?
    Sometimes, real fast is almost as good as real time.
    Just remember, Semper Gumbi - always be flexible!

  7. #6
    Just Joined!
    Join Date
    Mar 2011
    Posts
    9
    I don't whether our system offers this ccNUMA ability or not.
    But i think so. Only ich cann't see this ability on Opteron. From my measure i see no difference between bandwidths of two cores and it is unimportant where the two cores stand. But on nehalem.

  8. #7
    Just Joined!
    Join Date
    Mar 2011
    Posts
    9
    I hope it was my personal supercomputer. hahahaha.
    I make a measurment to controll whether the different memory strategies from intel, GNU and pgi do the right thing.
    Alternative mit time measurment can i do this right, but it isn't so nice.

  9. #8
    Linux Guru Rubberman's Avatar
    Join Date
    Apr 2009
    Location
    I can be found either 40 miles west of Chicago, in Chicago, or in a galaxy far, far away.
    Posts
    11,753
    As I said, most stand-alone SMP computers will not implement NUMA architecture features, which basically means that each CPU and/or core has its own memory, instead of sharing equally all memory with all cores. As noted, except in some cases, it requires specialized hardware, and definitely specialized operating system software. So, again the question is why do you need to do this "exercise"?
    Sometimes, real fast is almost as good as real time.
    Just remember, Semper Gumbi - always be flexible!

  10. #9
    Just Joined!
    Join Date
    Mar 2011
    Posts
    9
    OpenMP has a "task" construct. Two threads from different sockets alloc memorys. Normally these two fragment memorys are on thir own RAM. i.e. no remote. Now the two threads would like produce some tasks, they operate on the data that by their fatherthread alloced. But these tasks don't run immediately until their fathers finish. Now these tasks can run either on those cores on which tasks produced are or on strange cores.
    Obviously is first case the best. But some faktors influence those choices, for example, father thread run-time. And the compiler has an important effect too.

    This "exercise" is to find out how compilers act and which compiler has a better strategy.

  11. #10
    Linux Guru Rubberman's Avatar
    Join Date
    Apr 2009
    Location
    I can be found either 40 miles west of Chicago, in Chicago, or in a galaxy far, far away.
    Posts
    11,753
    Well, this is why you are in school, taking these classes, so you understand a) how it all works, and b) how to analyze these sort of issues in your work after you finish school, which is what they will (hopefully) be paying you the big buck$ for. Anyway, forum rules state that we aren't supposed to help with your homework/schoolwork except to point you in the general right direction, which hopefully we have done, at least somewhat. Since I have no ccNUMA system to work with, this has not been an issue for me. I have friends who are very versed on this technology, but they are not around to help - one passed away last year (he had his own Wave Tracer VLIW super computer), and the other, one of the principal architects of the x86 processor family at Intel, is retired and on a trip somewhere. So, remember that Google is your friend, and there is a fair amount of information in Wikipedia that may help get you going in the right direction.
    Sometimes, real fast is almost as good as real time.
    Just remember, Semper Gumbi - always be flexible!

Posting Permissions

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