Always use all threads in NumaTopology, cpuset respected when selecti…

…ng then, e.g. for affinity domains
RRZE-HPC · Sep 12, 2024 · 35f988b · 35f988b
1 parent b20cf61
commit 35f988b
Show file tree

Hide file tree

Showing 2 changed files with 21 additions and 14 deletions.
diff --git a/src/affinity.c b/src/affinity.c
@@ -527,7 +527,8 @@ static int affinity_addMemoryDomain(int nodeId, AffinityDomain* domain, int* hel
 {
     CpuTopology_t cputopo = get_cpuTopology();
     NumaTopology_t numatopo = get_numaTopology();
-    if (!domain)
+    int num_hwthreads = 0;
+    if ((nodeId < 0) || (!domain))
     {
         return -EINVAL;
     }
@@ -542,9 +543,15 @@ static int affinity_addMemoryDomain(int nodeId, AffinityDomain* domain, int* hel
             }
             for (int i = 0; i < numatopo->nodes[nodeId].numberOfProcessors; i++)
             {
-                domain->processorList[i] = (int)numatopo->nodes[nodeId].processors[i];
+                for (int j = 0; j < cputopo->numHWThreads; j++)
+                {
+                    if (cputopo->threadPool[j].apicId == numatopo->nodes[nodeId].processors[i] && cputopo->threadPool[j].inCpuSet == 1)
+                    {
+                        domain->processorList[num_hwthreads++] = (int)numatopo->nodes[nodeId].processors[i];
+                    }
+                }
             }
-            domain->numberOfProcessors = numatopo->nodes[nodeId].numberOfProcessors;
+            domain->numberOfProcessors = num_hwthreads;
             domain->numberOfCores = affinity_countSocketCores(domain->numberOfProcessors, domain->processorList, help);
             domain->tag = bformat("M%d", nodeId);
             DEBUG_PRINT(DEBUGLEV_DEVELOP, Affinity domain M%d: %d HW threads on %d cores, nodeId, domain->numberOfProcessors, domain->numberOfCores);

diff --git a/src/numa_proc.c b/src/numa_proc.c
@@ -372,17 +372,17 @@ int proc_numa_init(void)
         numa_info.nodes[i].id = id;
         nodeMeminfo(id, &numa_info.nodes[i].totalMemory, &numa_info.nodes[i].freeMemory);
         numa_info.nodes[i].numberOfProcessors = nodeProcessorList(id, &numa_info.nodes[i].processors);
-        int check = 0;
-        for (int j = 0; j < numa_info.nodes[i].numberOfProcessors; j++)
-        {
-            int id = numa_info.nodes[i].processors[j];
-            if (cpuid_topology.threadPool[id].inCpuSet == 1)
-            {
-                check++;
-            }
-        }
-        if (check < numa_info.nodes[i].numberOfProcessors)
-            numa_info.nodes[i].numberOfProcessors = check;
+/*        int check = 0;*/
+/*        for (int j = 0; j < numa_info.nodes[i].numberOfProcessors; j++)*/
+/*        {*/
+/*            int id = numa_info.nodes[i].processors[j];*/
+/*            if (cpuid_topology.threadPool[id].inCpuSet == 1)*/
+/*            {*/
+/*                numa_info.nodes[i].processors[check++] = numa_info.nodes[i].processors[j];*/
+/*            }*/
+/*        }*/
+/*        if (check < numa_info.nodes[i].numberOfProcessors)*/
+/*            numa_info.nodes[i].numberOfProcessors = check;*/
         numa_info.nodes[i].numberOfDistances = nodeDistanceList(id, numa_info.numberOfNodes, &numa_info.nodes[i].distances);
         if (numa_info.nodes[i].numberOfDistances == 0)
         {