Allow machine hierarchy expansion

[jlpeyton]

Summary

This fix allows the machine hierarchy to be expanded in case it needs to handle more threads. It adds a resize function to accomplish this.

[hfinkel]

I don't understand this patch; can you please explain more?

Is the idea that if, at runtime, someone adds more CPUs, the hierarchy will resize itself?

runtime/src/kmp_affinity.cpp

320

Why are you making this dynamic? How could this end up needing more than 7?

[tlwilmar]

Oversubscription. There's nothing to prevent the user from requesting an arbitrarily large number of threads.

[tlwilmar]

So, yes, this allows the hierarchy to resize itself in the oversubscription case.

[hfinkel]

In D9900#178569, @tlwilmar wrote:

Oversubscription. There's nothing to prevent the user from requesting an arbitrarily large number of threads.

But I thought that the machine hierarchy information represented the physical properties of the machine (cpus, cores, sockets, etc.). How does oversubscription factor into this? And how does that affect the number of levels?

[tlwilmar]

It does represent the machine hierarchy, when there is no oversubscription, but it is actually used as a way to map existing threads to the machine hierarchy for the purpose of performing a barrier. Adding a level when oversubscribed simply adds more tree structure for all the additional threads to communicate along. One additional level duplicates the existing hierarchy as a second child of a new root, doubling the number of threads that the barrier can handle.

[jlpeyton]

Added comments to clarify

Okay; please add a comment explaining this; the current definition of maxLevels says:
and this gives no indication at all of how this might need to be set to handle oversubscription. There are a few ...
but that does not really explain what is going on.

runtime/src/kmp_affinity.cpp

__kmp_affinity_cmp_Address_child_num(const void *a, const void *b)
     return 0;
 }
 
 /** A structure for holding machine-specific hierarchy info to be computed once at init. */
 class hierarchy_info {
 public:
     /** Typical levels are threads/core, cores/package or socket, packages/node, nodes/machine,
         etc.  We don't want to get specific with nomenclature */
-    static const kmp_uint32 maxLevels=7;
+    kmp_uint32 maxLevels;
 
     /** This is specifically the depth of the machine configuration hierarchy, in terms of the
         number of levels along the longest path from root to any leaf. It corresponds to the
         number of entries in numPerLevel if we exclude all but one trailing 1. */
     kmp_uint32 depth;
     kmp_uint32 base_num_threads;
     volatile kmp_int8 uninitialized; // 0=initialized, 1=uninitialized, 2=initialization in progress
+    volatile kmp_int8 resizing; // 0=not resizing, 1=resizing
 
     /** Level 0 corresponds to leaves. numPerLevel[i] is the number of children the parent of a
         node at level i has. For example, if we have a machine with 4 packages, 4 cores/package
         and 2 HT per core, then numPerLevel = {2, 4, 4, 1, 1}. All empty levels are set to 1. */
-    kmp_uint32 numPerLevel[maxLevels];
-    kmp_uint32 skipPerLevel[maxLevels];
+    kmp_uint32 *numPerLevel;
+    kmp_uint32 *skipPerLevel;
 
     void deriveLevels(AddrUnsPair *adr2os, int num_addrs) {
         int hier_depth = adr2os[0].first.depth;
         int level = 0;
         for (int i=hier_depth-1; i>=0; --i) {
             int max = -1;
             for (int j=0; j<num_addrs; ++j) {
                 int next = adr2os[j].first.childNums[i];
                 if (next > max) max = next;
             }
             numPerLevel[level] = max+1;
             ++level;
         }
     }
 
-    hierarchy_info() : depth(1), uninitialized(1) {}
+    hierarchy_info() : maxLevels(7), depth(1), uninitialized(1), resizing(0) {}
+
+    // TO FIX: This destructor causes a segfault in the library at shutdown.
+    //~hierarchy_info() { if (!uninitialized && numPerLevel) __kmp_free(numPerLevel); }
+
     void init(AddrUnsPair *adr2os, int num_addrs)
     {
         kmp_int8 bool_result = KMP_COMPARE_AND_STORE_ACQ8(&uninitialized, 1, 2);
         if (bool_result == 0) { // Wait for initialization
             while (TCR_1(uninitialized) != 0) KMP_CPU_PAUSE();
             return;
         }
         KMP_DEBUG_ASSERT(bool_result==1);
 
-        /* Added explicit initialization of the depth here to prevent usage of dirty value
+        /* Added explicit initialization of the data fields here to prevent usage of dirty value
            observed when static library is re-initialized multiple times (e.g. when
            non-OpenMP thread repeatedly launches/joins thread that uses OpenMP). */
         depth = 1;
+        resizing = 0;
+        maxLevels = 7;
+        numPerLevel = (kmp_uint32 *)__kmp_allocate(maxLevels*2*sizeof(kmp_uint32));
+        skipPerLevel = &(numPerLevel[maxLevels]);
         for (kmp_uint32 i=0; i<maxLevels; ++i) { // init numPerLevel[*] to 1 item per level
             numPerLevel[i] = 1;
             skipPerLevel[i] = 1;
         }
 
         // Sort table by physical ID
         if (adr2os) {
             qsort(adr2os, num_addrs, sizeof(*adr2os), __kmp_affinity_cmp_Address_labels);
void init(AddrUnsPair *adr2os, int num_addrs)
             skipPerLevel[i] = numPerLevel[i-1] * skipPerLevel[i-1];
         // Fill in hierarchy in the case of oversubscription
         for (kmp_uint32 i=depth; i<maxLevels; ++i)
             skipPerLevel[i] = 2*skipPerLevel[i-1];
 
         uninitialized = 0; // One writer
 
     }
+
+    void resize(kmp_uint32 nproc)
+    {
+        kmp_int8 bool_result = KMP_COMPARE_AND_STORE_ACQ8(&resizing, 0, 1);
+        if (bool_result == 0) { // Someone else is resizing
+            while (TCR_1(resizing) != 0) KMP_CPU_PAUSE();
+            return;
+        }
+        KMP_DEBUG_ASSERT(bool_result!=0);
+        KMP_DEBUG_ASSERT(nproc > base_num_threads);
+
+        // Calculate new max_levels
+        kmp_uint32 old_sz = skipPerLevel[depth-1];
+        kmp_uint32 incs = 0, old_maxLevels= maxLevels;
+        while (nproc > old_sz) {
+            old_sz *=2;
+            incs++;
+        }
+        maxLevels += incs;
+
+        // Resize arrays
+        kmp_uint32 *old_numPerLevel = numPerLevel;
+        kmp_uint32 *old_skipPerLevel = skipPerLevel;
+        numPerLevel = skipPerLevel = NULL;
+        numPerLevel = (kmp_uint32 *)__kmp_allocate(maxLevels*2*sizeof(kmp_uint32));
+        skipPerLevel = &(numPerLevel[maxLevels]);
+
+        // Copy old elements from old arrays
+        for (kmp_uint32 i=0; i<old_maxLevels; ++i) { // init numPerLevel[*] to 1 item per level
+            numPerLevel[i] = old_numPerLevel[i];
+            skipPerLevel[i] = old_skipPerLevel[i];
+        }
+
+        // Init new elements in arrays to 1
+        for (kmp_uint32 i=old_maxLevels; i<maxLevels; ++i) { // init numPerLevel[*] to 1 item per level
+            numPerLevel[i] = 1;
+            skipPerLevel[i] = 1;
+        }
+
+        // Free old arrays
+        __kmp_free(old_numPerLevel);
+
+        // Fill in oversubscription levels of hierarchy
+        for (kmp_uint32 i=old_maxLevels; i<maxLevels; ++i)
+            skipPerLevel[i] = 2*skipPerLevel[i-1];
+
+        base_num_threads = nproc;
+        resizing = 0; // One writer
+
+    }
 };
 
 static hierarchy_info machine_hierarchy;
 
 void __kmp_get_hierarchy(kmp_uint32 nproc, kmp_bstate_t *thr_bar) {
     kmp_uint32 depth;
     // The test below is true if affinity is available, but set to "none". Need to init on first use of hierarchical barrier.
     if (TCR_1(machine_hierarchy.uninitialized))
         machine_hierarchy.init(NULL, nproc);
+    // Adjust the hierarchy in case num threads exceeds original 
+    if (nproc > machine_hierarchy.base_num_threads)
+        machine_hierarchy.resize(nproc);
 
     depth = machine_hierarchy.depth;
     KMP_DEBUG_ASSERT(depth > 0);
-    // The loop below adjusts the depth in the case of oversubscription
-    while (nproc > machine_hierarchy.skipPerLevel[depth-1] && depth<machine_hierarchy.maxLevels-1)
+    // The loop below adjusts the depth in the case of a resize
+    while (nproc > machine_hierarchy.skipPerLevel[depth-1])
         depth++;
 
     thr_bar->depth = depth;
     thr_bar->base_leaf_kids = (kmp_uint8)machine_hierarchy.numPerLevel[0]-1;
     thr_bar->skip_per_level = machine_hierarchy.skipPerLevel;
 }
 
 //