Move memory pages per-thread in NUMA architecture

问题

i have 2 questions in one:

(i) Suppose thread X is running at CPU Y. Is it possible to use the syscalls migrate_pages - or even better move_pages (or their libnuma wrapper) - to move the pages associated with X to the node in which Y is connected?

This question arrises because first argument of both syscalls is PID (and i need a per-thread approach for some researching i'm doing)

(ii) in the case of positive answer for (i), how can i get all the pages used by some thread? My aim is, move the page(s) that contains array M[] for exemple...how to "link" data structures with their memory pages, for the sake of using the syscalls above?

An extra information: i'm using C with pthreads. Thanks in advance !

回答1:

You want to use the higher level libnuma interfaces instead of the low level system calls.

The libnuma library offers a simple programming interface to the NUMA (Non Uniform Memory Access) policy supported by the Linux kernel. On a NUMA architecture some memory areas have different latency or bandwidth than others.

Available policies are page interleaving (i.e., allocate in a round-robin fashion from all, or a subset, of the nodes on the system), preferred node allocation (i.e., preferably allocate on a particular node), local allocation (i.e., allocate on the node on which the task is currently executing), or allocation only on specific nodes (i.e., allocate on some subset of the available nodes). It is also possible to bind tasks to specific nodes.

The man pages for the low level numa_* system calls warn you away from using them:

Link with -lnuma to get the system call definitions. libnuma and the required <numaif.h> header are available in the numactl package.

However, applications should not use these system calls directly. Instead, the higher level interface provided by the numa(3) functions in the numactl package is recommended. The numactl package is available at <ftp://oss.sgi.com/www/projects/libnuma/download/>. The package is also included in some Linux distributions. Some distributions include the development library and header in the separate numactl-devel package.

回答2:

Here's the code I use for pinning a thread to a single CPU and moving the stack to the corresponding NUMA node (slightly adapted to remove some constants defined elsewhere). Note that I first create the thread normally, and then call the SetAffinityAndRelocateStack() below from within the thread. I think this is much better then trying to create your own stack, since stacks have special support for growing in case the bottom is reached.

The code can also be adapted to operate on the newly created thread from outside, but this could give rise to race conditions (e.g. if the thread performs I/O into its stack), so I wouldn't recommend it.

void* PreFaultStack()
{
    const size_t NUM_PAGES_TO_PRE_FAULT = 50;
    const size_t size = NUM_PAGES_TO_PRE_FAULT * numa_pagesize();
    void *allocaBase = alloca(size);
    memset(allocaBase, 0, size);
    return allocaBase;
}

void SetAffinityAndRelocateStack(int cpuNum)
{
    assert(-1 != cpuNum);
    cpu_set_t cpuset;
    CPU_ZERO(&cpuset);
    CPU_SET(cpuNum, &cpuset);
    const int rc = pthread_setaffinity_np(pthread_self(), sizeof(cpu_set_t), &cpuset);
    assert(0 == rc);

    pthread_attr_t attr;
    void *stackAddr = nullptr;
    size_t stackSize = 0;
    if ((0 != pthread_getattr_np(pthread_self(), &attr)) || (0 != pthread_attr_getstack(&attr, &stackAddr, &stackSize))) {
        assert(false);
    }

    const unsigned long nodeMask = 1UL << numa_node_of_cpu(cpuNum);
    const auto bindRc = mbind(stackAddr, stackSize, MPOL_BIND, &nodeMask, sizeof(nodeMask), MPOL_MF_MOVE | MPOL_MF_STRICT);
    assert(0 == bindRc);

    PreFaultStack();
    // TODO: Also lock the stack with mlock() to guarantee it stays resident in RAM
    return;
}

来源：https://stackoverflow.com/questions/18009230/move-memory-pages-per-thread-in-numa-architecture