make topology signature include a compressed list of available PUs

When mpirun collects topologies it first asks for signatures and then
only asks a host to send its topologif the signature is something new.
But with heterogeneous cgroups the signature doesn't have enough data
to notice differences that matter.

An example of the previous signature on one of our machines is
    2N:2S:22L3:22L2:44L1:44C:176H:ppc64le:le
which is just a count of how many numa nodes, sockets, cores, PUs etc
the host has.  If I create a small cgroup with just 4 PUs from one of
the cores, an example of the new signature is
    1N:1S:1L3:1L2:1L1:1C:4H:ppc64le:le

But if I did that on two different machines and used different cores
on each, the signature would look the same. One way to distinguish
such signatures is to also list the available PUs by os_index (I might
be okay with logical indexes if it was logical under a WHOLE-SYSTEM
topology, but in the non-WHOLE-SYSTEM topologies we're using  I don't
think the logical indexes would be good enough).

This list could be large so I've included a couple different
compressions of it and use whichever comes out shorter.

The first is a comma separated list that compresses simple ranges like
10-15 and also detects patterns analogous to MPI_Type_vector(), eg
    {16,17,18,19,20,21,22,23} = 16-23
    {1,2,3,4,5,  11,12,13,14,15,  21,22,23,24,25} = 1-5+10*3
    {2,3, 10,11, 18,19, 26,27,
     102,103, 110,111, 118,119,
     200,201,202,203,204,205,206,207,208 } = 2-3+8*4,102-103+8*3,200-208
    {1,3,6,7,9,11,12} = 1,3,6,7,9,11,12

The second compression is a hex string containing the indexes, and
further shrunk by noticing if the same char is repeated 5+ times, so
    {16,17,18,19,20,21,22,23} = 0000ff
    {1,2,3,4,5,  11,12,13,14,15,  21,22,23,24,25} = 7c1f07c
    {2,3, 10,11, 18,19, 26,27,
     102,103, 110,111, 118,119,
     200,201,202,203,204,205,206,207,208 } = 30303030*18,303030*20,ff8
    {1,3,6,7,9,11,12} = 5358

So final signature strings end up things like

example with PUs 0-175:
    2N:2S:22L3:22L2:44L1:44C:176H:HXf*44:ppc64le:le
here "f*44" won the compression contest against "0-175"

example with PUs 12-15:
    1N:1S:1L3:1L2:1L1:1C:4H:HX000f:ppc64le:le
here "000f" won against "12-15"

example with every other block of 4 PUs:
    2N:2S:22L3:22L2:22L1:22C:88H:HL0-3+8*22:ppc64le:le
here "0-3+8*22" won against "f0f0f0f0f0f0f0f0f0f0f0f0f0f0f0f0f0f0f0f0f0f"

Signed-off-by: Mark Allen <[email protected]>

Author: markalle

opal/mca/hwloc/base/hwloc_base_util.c

@@ -2251,6 +2251,254 @@ int opal_hwloc_get_sorted_numa_list(hwloc_topology_t topo, char* device_name, op
     return OPAL_ERR_NOT_FOUND;
 }
 
+static int
+mycompare(const void* inta, const void* intb)
+{
+    if (*(int*)inta < *(int*)intb) { return -1; }
+    if (*(int*)inta > *(int*)intb) { return 1; }
+    return 0;
+}
+
+// Patterns to compress adequately:
+// 1. basic 0,1,2,3 to 0-3
+// 2. repeated ranges of same len at same offset, eg 0,1,2 10,11,12 20,21,22 = 0,1,2+10*3
+// note: assumes input is already sorted
+static char*
+mycompress1(int *a, int n)
+{
+    char *str;
+    int i, maxlen, len;
+
+    maxlen = 128;
+    len = 0;
+    str = malloc(maxlen);
+    if (!str) { return NULL; }
+    str[0] = 0;
+
+    // start with "L" followed by the list
+    str[len++] = 'L';
+    str[len] = 0;
+
+    int pattern_blocksize = 0; // contiguous elements per block
+    int pattern_blockcount = 0; // number of blocks
+    int pattern_stride = 0;
+    int pattern_first_block_starting_offset = -1;
+    int pattern_prev_block_starting_offset = -1;
+    int pattern_nelements = 0;
+    int is_start_of_new_pattern = 0;
+    for (i=0; i<n; ++i) {
+        // see if this element a[i] is the start of a new pattern or
+        // if it's adding to the previous, we may increment i further
+        // processing multiple elements into the previous pattern
+        // only update the pattern_* data for the previous pattern
+        // (if we're starting a new pattern, we put it into pattern_*,
+        // later, after printing what was already in pattern_*)
+
+        // start by supposing it is a new pattern, change that decision
+        // if we process it into the existing pattern_* below
+        is_start_of_new_pattern = 1;
+
+        if (pattern_blocksize > 0) {
+            // detect [pppx]
+            if (pattern_blockcount == 1 &&
+                a[i] == pattern_prev_block_starting_offset + pattern_blocksize)
+            {
+                ++pattern_blocksize;
+                ++pattern_nelements;
+                is_start_of_new_pattern = 0;
+            } else {
+                // detect [pppp  xxxx]   or   [pppp  pppp  xxxx]   etc
+                if (pattern_blockcount == 1 ||
+                   (pattern_blockcount > 1 &&
+                    (a[i] == pattern_prev_block_starting_offset + pattern_stride)))
+                {
+                    // this a[i] starts at the right offset, now does it contain
+                    // pattern_blocksize contiguous entries?
+                    int j, has_enough_entries = 1;
+                    for (j=1; j<pattern_blocksize; ++j) {
+                        if (i+j >= n) { has_enough_entries = 0; }
+                        else if (a[i+j] != a[i] + j) { has_enough_entries = 0; }
+                    }
+                    if (has_enough_entries) {
+                        pattern_stride = a[i] - pattern_prev_block_starting_offset;
+                        pattern_prev_block_starting_offset = a[i];
+                        ++pattern_blockcount;
+                        pattern_nelements += pattern_blocksize;
+                        is_start_of_new_pattern = 0;
+                        i = i+j-1;
+                    }
+                }
+            }
+        }
+
+        // if started_new_pattern || i is the last element
+        //   if previous pattern exists
+        //     print previous pattern
+        //   record new pattern
+        if (is_start_of_new_pattern || i==n-1) {
+            if (pattern_blocksize > 0) {
+                // print previous pattern
+                // make sure there's room for 4 ints plus some punctuation
+                // being conservative in supposing all ints used are <8 chars
+                if (len + 40 >= maxlen) {
+                    maxlen *= 1.15;
+                    maxlen += 128;
+                    str = realloc(str, maxlen);
+                    if (!str) { return NULL; }
+                }
+                // If the string isn't 'long enough' for compressed notation to be worthwhile:
+                if (!(pattern_nelements > 4 || (pattern_nelements >= 2 && pattern_blockcount == 1))) {
+                    int j, k;
+                    for (j=0; j<pattern_blockcount; ++j) {
+                        for (k=0; k<pattern_blocksize; ++k) {
+                            if (len != 1) { // the string starts with "L" for list
+                                sprintf(&str[len], ",");
+                                ++len;
+                            }
+                            sprintf(&str[len], "%d", pattern_first_block_starting_offset + j*pattern_stride + k);
+                            len = strlen(str);
+                        }
+                    }
+                }
+                else {
+                    if (len != 1) { // the string starts with "L" for list
+                        sprintf(&str[len], ",");
+                        ++len;
+                    }
+                    if (pattern_blockcount > 1) {
+                        sprintf(&str[len], "%d-%d+%d*%d",
+                            pattern_first_block_starting_offset,
+                            pattern_first_block_starting_offset + pattern_blocksize - 1,
+                            pattern_stride, pattern_blockcount);
+                    } else {
+                        sprintf(&str[len], "%d-%d",
+                            pattern_first_block_starting_offset,
+                            pattern_first_block_starting_offset + pattern_blocksize - 1);
+                    }
+                    len = strlen(str);
+
+                }
+            }
+
+            // record new pattern
+            pattern_first_block_starting_offset = a[i];
+            pattern_prev_block_starting_offset = a[i];
+            pattern_blocksize = 1;
+            pattern_blockcount = 1;
+            pattern_stride = 0;
+            pattern_nelements = 1;
+        }
+        // The overall logic till this point is
+        // loop i over some of a[], and either
+        // 1. put a[i] in the previous pattern (and print if it's end of a[])
+        // 2. put a[i] in a new pattern and print the previous pattern
+        // This leaves out printing in the case of a[i] being a new pattern
+        // and also being the end of a[]
+        if (is_start_of_new_pattern && i==n-1) {
+            if (len != 0) {
+                sprintf(&str[len], ",");
+                ++len;
+            }
+            sprintf(&str[len], "%d", pattern_first_block_starting_offset);
+            len = strlen(str);
+        }
+    }   
+            
+    return str;
+}
+// This one just makes a hex string of the available bits
+// update: and does a tiny bit of trivial compression of adjacent
+// hex characters
+// note: assumes input is already sorted
+static char*
+mycompress2(int *a, int n)
+{
+    unsigned char *hex_values;
+    char *str, prev_char, *p;
+    int i, len, max, count;
+
+    max = a[n-1];
+    hex_values = malloc((max / 4 + 4) * sizeof(unsigned char));
+    str = malloc(max/4 + 32);
+    if (!str || !hex_values) { return NULL; }
+
+    // start with "X" followed by the hex bitmask
+    len = 0;
+    str[len++] = 'X';
+
+    for (i=0; i<=max/4; ++i) {
+        hex_values[i] = 0;
+    }
+    for (i=0; i<n; ++i) {
+        int bit = a[i];
+
+        hex_values[bit/4] |= (unsigned char)(8>>(bit % 4));
+        // eg if bit = 11 then hex_values[2] |= 8>>3, eg [0000;0000;0001]
+    }
+
+    for (i=0; i<=max/4; ++i) {
+        sprintf(&str[len], "%x", hex_values[i]);
+        ++len;
+    }
+    free(hex_values);
+    str[len++] = 0;
+
+    // Add a trivial compression of adjacent chars that are the same so 0000 = 0*4,
+    p = str;
+    count = 0;
+    prev_char = (char)0;
+    for (i=0; i<len+1; ++i) {
+        if (i==len || str[i] != prev_char) {
+            // new section, so print pattern for previously stored stuff,
+            // then store new section
+            if (prev_char != (char)0) {
+                if (count <= 4) {
+                    int j;
+                    for (j=0; j<count; ++j) {
+                        *p = prev_char;
+                        ++p;
+                    }   
+                } else {
+                    sprintf(p, "%c*%d", prev_char, count);
+                    p += strlen(p);
+                    if (i < len-1) {
+                        *p = ',';
+                        ++p;
+                    }
+                }   
+            }   
+            prev_char = str[i];
+            count = 1;
+        } else {
+            // new char is part of previous section
+            ++count;
+        }   
+    }
+    *p = 0;
+
+    return str;
+}
+
+// Use two completely different compressions and pick the shorter
+static char*
+mycompress(int *a, int n)
+{
+    char *str1, *str2;
+
+    str1 = mycompress1(a, n);
+    str2 = mycompress2(a, n);
+
+    if (strlen(str2) < strlen(str1)) {
+        free(str1);
+        return str2;
+    }
+    free(str2);
+    return str1;
+}
+
+// The only parsing of this string I can find is a strrchr(str,":")
+// to evaluate the endianness of the nodes, so we have to let that
+// remain as the last entry 
 char* opal_hwloc_base_get_topo_signature(hwloc_topology_t topo)
 {
     int nnuma, nsocket, nl3, nl2, nl1, ncore, nhwt;
@@ -2288,8 +2536,47 @@ char* opal_hwloc_base_get_topo_signature(hwloc_topology_t topo)
     endian = "unknown";
 #endif
 
-    opal_asprintf(&sig, "%dN:%dS:%dL3:%dL2:%dL1:%dC:%dH:%s:%s",
-             nnuma, nsocket, nl3, nl2, nl1, ncore, nhwt, arch, endian);
+// The signature used to just contain number of nodes, sockets, cores,
+// hardware threads, etc.  With heterogeneous cgroups it's not hard to
+// have the counts come out the same even though the topologies of the
+// available elements are different enough to matter
+//
+// Example hardware:
+//    [..../..../..../....][..../..../..../....]
+//     0    4    8    12    16   20   24   28
+// cgset -r cpuset.cpus=10,11,14,15 mycgroup1
+// cgset -r cpuset.cpus=26,27,30,31 mycgroup2
+// the above cgroups would leave only these hardware threads active:
+//     mycgroup1: [~~~~/~~~~/~~HH/~~HH][~~~~/~~~~/~~~~/~~~~]
+//     mycgroup2: [~~~~/~~~~/~~~~/~~~~][~~~~/~~~~/~~HH/~~HH]
+// The non-WHOLE-SYSTEM topology would only contain the availabe part
+// of the tree and each host would report as 1 socket, 2 cores, etc
+//
+// I think a compressed list of the os_indexes in the available mask
+// should be a modest-sized addition to the string that would suffice
+// to make the signatures differ for different cgroups
+
+    int npu;
+    int *avail_pus;
+    char *avail_pus_string;
+    obj = NULL;
+    npu = hwloc_get_nbobjs_by_type(topo, HWLOC_OBJ_PU);
+    avail_pus = malloc(npu * sizeof(int));
+    if (!avail_pus) { return strdup("malloc failed"); }
+    npu = 0;
+    while ((obj = hwloc_get_next_obj_by_type(topo, HWLOC_OBJ_PU, obj)) != NULL) {
+        avail_pus[npu++] = obj->os_index;
+    }
+    qsort(avail_pus, npu, sizeof(int), &mycompare);
+
+    avail_pus_string = mycompress(avail_pus, npu);
+    opal_asprintf(&sig, "%dN:%dS:%dL3:%dL2:%dL1:%dC:%dH:H%s:%s:%s",
+             nnuma, nsocket, nl3, nl2, nl1, ncore, nhwt,
+             avail_pus_string?avail_pus_string:"", arch, endian);
+
+    free(avail_pus);
+    if (avail_pus_string) { free(avail_pus_string); }
+
     return sig;
 }