diff --git a/CMakeLists.txt b/CMakeLists.txt
index ac3e9090336d9..715adb13e3120 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -83,6 +83,7 @@ option(LLAMA_BUILD_SERVER   "llama: build server example" ${LLAMA_STANDALONE})
 
 # 3rd party libs
 option(LLAMA_CURL       "llama: use libcurl to download model from an URL" ON)
+option(LLAMA_NUMA       "llama: use libnuma to get memory policy of the llama-bench" ON)
 option(LLAMA_LLGUIDANCE "llama-common: include LLGuidance library for structured output in common utils" OFF)
 
 # Required for relocatable CMake package
diff --git a/src/CMakeLists.txt b/src/CMakeLists.txt
index d4bf37b1cf3e5..17079c7f25c47 100644
--- a/src/CMakeLists.txt
+++ b/src/CMakeLists.txt
@@ -39,6 +39,19 @@ target_compile_features   (llama PRIVATE cxx_std_17) # don't bump
 
 target_link_libraries(llama PUBLIC ggml)
 
+if (LLAMA_NUMA)
+    find_library(NUMA_LIB numa)
+    check_include_file_cxx("numa.h" HAVE_NUMA_HEADERS)
+    check_include_file_cxx("numaif.h" HAVE_NUMAIF_HEADERS)
+    if (HAVE_NUMA_HEADERS AND HAVE_NUMAIF_HEADERS AND NUMA_LIB)
+        target_compile_definitions(llama PRIVATE USE_LIBNUMA)
+        target_link_libraries(llama PRIVATE numa)
+        message(STATUS "libnuma found, page cache will be moved to the local node using mbind() syscall. Disable with LLAMA_NUMA=OFF")
+    else()
+        message(STATUS "Warning: NUMA headers not found - consider disabling this Warning with LLAMA_NUMA=OFF")
+    endif()
+endif()
+
 if (BUILD_SHARED_LIBS)
     set_target_properties(llama PROPERTIES POSITION_INDEPENDENT_CODE ON)
     target_compile_definitions(llama PRIVATE LLAMA_BUILD)
diff --git a/src/llama-mmap.cpp b/src/llama-mmap.cpp
index 9da97f1bc5057..44dd861423108 100644
--- a/src/llama-mmap.cpp
+++ b/src/llama-mmap.cpp
@@ -10,6 +10,12 @@
 #include <cerrno>
 #include <algorithm>
 
+#ifdef USE_LIBNUMA
+    #include <numa.h>
+    #include <numaif.h>
+    #include <sched.h>
+#endif
+
 #ifdef __has_include
     #if __has_include(<unistd.h>)
         #include <unistd.h>
@@ -273,6 +279,27 @@ struct llama_mmap::impl {
 #ifdef _POSIX_MAPPED_FILES
     std::vector<std::pair<size_t, size_t>> mapped_fragments;
 
+#ifdef USE_LIBNUMA
+    static void move_pages(void *addr, size_t size) {
+        int cpu, ret;
+        struct bitmask *nodemask = numa_allocate_nodemask();
+
+        /* Get memory policy of the calling thread. */
+        ret = get_mempolicy(nullptr, nodemask->maskp, nodemask->size, nullptr, 0);
+        if (ret || numa_bitmask_weight(nodemask) == 0) {
+            cpu = sched_getcpu();
+            if (cpu >= 0) {
+                numa_bitmask_clearall(nodemask);
+                numa_bitmask_setbit(nodemask, numa_node_of_cpu(cpu));
+            }
+        }
+        if (numa_bitmask_weight(nodemask) == 1) {
+            mbind(addr, size, MPOL_BIND, nodemask->maskp, nodemask->size, MPOL_MF_MOVE);
+        }
+        numa_free_nodemask(nodemask);
+    }
+#endif
+
     impl(struct llama_file * file, size_t prefetch, bool numa) {
         size = file->size();
         int fd = file->file_id();
@@ -291,6 +318,17 @@ struct llama_mmap::impl {
         }
 
         if (prefetch > 0) {
+#ifdef USE_LIBNUMA
+            /*
+             * Given that we already pre-fault all memory when prefetch > 0, it is
+             * necessary to move any page cache pages that might have been
+             * instantiated during previous runs on different NUMA nodes. This call
+             * to move_pages() ensures that all memory-mapped pages are relocated
+             * according to the calling thread's memory policy or the CPU on which
+             * it is running.
+             */
+            move_pages(addr, file->size());
+#endif
             if (posix_madvise(addr, std::min(file->size(), prefetch), POSIX_MADV_WILLNEED)) {
                 LLAMA_LOG_WARN("warning: posix_madvise(.., POSIX_MADV_WILLNEED) failed: %s\n",
                         strerror(errno));