diff --git a/include/mimalloc-types.h b/include/mimalloc-types.h
index 0208d5c7..99b6b22b 100644
--- a/include/mimalloc-types.h
+++ b/include/mimalloc-types.h
@@ -413,7 +413,6 @@ typedef struct mi_segments_tld_s {
 // OS thread local data
 typedef struct mi_os_tld_s {
   size_t              region_idx;   // start point for next allocation
-  int                 numa_node;    // numa node associated with this thread
   mi_stats_t*         stats;        // points to tld stats
 } mi_os_tld_t;
 
diff --git a/src/init.c b/src/init.c
index 166ca451..ed15aeba 100644
--- a/src/init.c
+++ b/src/init.c
@@ -99,8 +99,8 @@ static mi_tld_t tld_main = {
   0, false,
   &_mi_heap_main,
   { { NULL, NULL }, {NULL ,NULL}, 0, 0, 0, 0, 0, 0, NULL, tld_main_stats }, // segments
-  { 0, -1, tld_main_stats },   // os
-  { MI_STATS_NULL }            // stats
+  { 0, tld_main_stats },   // os
+  { MI_STATS_NULL }        // stats
 };
 
 mi_heap_t _mi_heap_main = {
@@ -218,7 +218,6 @@ static bool _mi_heap_init(void) {
     memset(tld, 0, sizeof(*tld));
     tld->heap_backing = heap;
     tld->segments.stats = &tld->stats;
-    tld->os.numa_node = -1;
     tld->os.stats = &tld->stats;
     _mi_heap_default = heap;
   }
diff --git a/src/os.c b/src/os.c
index c41d028f..8e1b3e91 100644
--- a/src/os.c
+++ b/src/os.c
@@ -903,29 +903,21 @@ static int mi_os_numa_node_countx(void) {
   GetNumaHighestNodeNumber(&numa_max);
   return (int)(numa_max + 1);
 }
-#elif MI_HAS_NUMA
+#elif defined(__linux__)
 #include <dirent.h>
 #include <stdlib.h>
-#include <numaif.h>
+#include <sys/syscall.h>
+
 static int mi_os_numa_nodex(void) {
-  #define MI_NUMA_NODE_SLOW  // too slow, so cache it
-  // TODO: perhaps use RDTSCP instruction on x64? 
-  // see <https://stackoverflow.com/questions/16862620/numa-get-current-node-core>
-  #define MI_MAX_MASK (4)          // support at most 256 nodes
-  unsigned long mask[MI_MAX_MASK];
-  memset(mask,0,MI_MAX_MASK*sizeof(long));
-  int mode = 0;
-  long err = get_mempolicy(&mode, mask, MI_MAX_MASK*sizeof(long)*8, NULL, 0 /* thread policy */);
+#ifdef SYS_getcpu
+  unsigned node = 0;
+  unsigned ncpu = 0;
+  int err = syscall(SYS_getcpu, &ncpu, &node, NULL);
   if (err != 0) return 0;
-  // find the lowest bit that is set
-  for(int i = 0; i < MI_MAX_MASK; i++) {
-    for(int j = 0; j < (int)(sizeof(long)*8); j++) {
-      if ((mask[i] & (1UL << j)) != 0) {
-        return (i*sizeof(long)*8 + j);
-      }
-    }
-  }
-	return 0;
+  return (int)node;
+#else
+  return 0;
+#endif
 }
 
 static int mi_os_numa_node_countx(void) {
@@ -968,21 +960,8 @@ int _mi_os_numa_node_count(void) {
 }
 
 int _mi_os_numa_node(mi_os_tld_t* tld) {
-  int numa_node;
-#ifndef MI_NUMA_NODE_SLOW
   UNUSED(tld);
-  numa_node = mi_os_numa_nodex();
-#else
-  if (mi_unlikely(tld->numa_node < 0)) {
-    // Cache the NUMA node of the thread if the call is slow.
-    // This may not be correct as threads can migrate to another cpu on
-    // another node -- however, for memory allocation this just means we keep
-    // using the same 'node id' for its allocations; new OS allocations
-    // naturally come from the actual node so in practice this may be fine.
-    tld->numa_node = mi_os_numa_nodex(); 
-  }
-  numa_node = tld->numa_node;
-#endif
+  int numa_node = mi_os_numa_nodex();
   // never more than the node count and >= 0
   int numa_count = _mi_os_numa_node_count();
   if (numa_node >= numa_count) { numa_node = numa_node % numa_count; }