diff --git a/include/mimalloc-types.h b/include/mimalloc-types.h
index 9beeb195..67ad8516 100644
--- a/include/mimalloc-types.h
+++ b/include/mimalloc-types.h
@@ -384,6 +384,7 @@ 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
   mi_stats_t*         stats;        // points to tld stats
 } mi_os_tld_t;
 
diff --git a/src/init.c b/src/init.c
index 0261e26b..77ce4aad 100644
--- a/src/init.c
+++ b/src/init.c
@@ -97,7 +97,7 @@ static mi_tld_t tld_main = {
   0,
   &_mi_heap_main,
   { { NULL, NULL }, {NULL ,NULL}, 0, 0, 0, 0, 0, 0, NULL, tld_main_stats }, // segments
-  { tld_main_stats },          // os
+  { 0, tld_main_stats },       // os
   { MI_STATS_NULL }            // stats
 };
 
diff --git a/src/memory.c b/src/memory.c
index ccd810b3..cf341105 100644
--- a/src/memory.c
+++ b/src/memory.c
@@ -79,7 +79,6 @@ typedef struct mem_region_s {
 static mem_region_t regions[MI_REGION_MAX];
 
 static volatile size_t regions_count = 0;        // allocated regions
-static volatile uintptr_t region_next_idx = 0;   // good place to start searching
 
 
 /* ----------------------------------------------------------------------------
@@ -180,7 +179,6 @@ static bool mi_region_commit_blocks(mem_region_t* region, size_t idx, size_t bit
   }
 
   // and return the allocation
-  mi_atomic_write(&region_next_idx,idx);  // next search from here
   *p  = blocks_start;
   *id = (idx*MI_REGION_MAP_BITS) + bitidx;
   return true;
@@ -267,7 +265,7 @@ void* _mi_mem_alloc_aligned(size_t size, size_t alignment, bool commit, size_t*
   // find a range of free blocks
   void* p = NULL;
   size_t count = mi_atomic_read(&regions_count);
-  size_t idx = mi_atomic_read(&region_next_idx);
+  size_t idx = tld->region_idx; // start index is per-thread to reduce contention
   for (size_t visited = 0; visited < count; visited++, idx++) {
     if (idx >= count) idx = 0;  // wrap around
     if (!mi_region_try_alloc_blocks(idx, blocks, size, commit, &p, id, tld)) return NULL; // error
@@ -286,6 +284,9 @@ void* _mi_mem_alloc_aligned(size_t size, size_t alignment, bool commit, size_t*
     // we could not find a place to allocate, fall back to the os directly
     p = _mi_os_alloc_aligned(size, alignment, commit, tld);
   }
+  else {
+    tld->region_idx = idx;  // next start of search
+  }
 
   mi_assert_internal( p == NULL || (uintptr_t)p % alignment == 0);
   return p;