ggml : add NUMA support (#1556)

* detect NUMA systems and pin work threads to nodes (linux)

* disable mmap prefetch/readahead for NUMA systems

* avoid sending finalize op to thread pool if it does nothing

* silence robot

* fix args

* make --numa a param

* recommendation that n_nodes evenly divide n_threads did not warrant such aggressive enforcement

* lower synchronization overhead

* statically allocate

* move numa state to g_state

* add description for --numa

* ggml : minor style changes

* ggml : minor style + try fix sanitizer build

* llama : allow to initialize backend with NUMA support

* llama : avoid ggml include in llama-util.h

* ggml : style / formatting

* ggml : fix handling of ops with n_threads > n_tasks > 1

* server : utilize numa parameter

---------

Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>

1 files changed, 19 insertions, 5 deletions

diff --git a/llama-util.h b/llama-util.h
index 4f8a429..042ebe4 100644
--- a/llama-util.h
+++ b/llama-util.h
@@ -172,12 +172,14 @@ struct llama_mmap {
 #ifdef _POSIX_MAPPED_FILES
     static constexpr bool SUPPORTED = true;
 
-    llama_mmap(struct llama_file * file, size_t prefetch = (size_t) -1 /* -1 = max value */) {
+    llama_mmap(struct llama_file * file, size_t prefetch = (size_t) -1 /* -1 = max value */, bool numa = false) {
         size = file->size;
         int fd = fileno(file->fp);
         int flags = MAP_SHARED;
+        // prefetch/readahead impairs performance on NUMA systems
+        if (numa) { prefetch = 0; }
 #ifdef __linux__
-        flags |= MAP_POPULATE;
+        if (prefetch) { flags |= MAP_POPULATE; }
 #endif
         addr = mmap(NULL, file->size, PROT_READ, flags, fd, 0);
         if (addr == MAP_FAILED) {
@@ -191,6 +193,14 @@ struct llama_mmap {
                         strerror(errno));
             }
         }
+        if (numa) {
+            // advise the kernel not to use readahead
+            // (because the next page might not belong on the same node)
+            if (madvise(addr, file->size, MADV_RANDOM)) {
+                fprintf(stderr, "warning: madvise(.., MADV_RANDOM) failed: %s\n",
+                        strerror(errno));
+            }
+        }
     }
 
     ~llama_mmap() {
@@ -199,7 +209,9 @@ struct llama_mmap {
 #elif defined(_WIN32)
     static constexpr bool SUPPORTED = true;
 
-    llama_mmap(struct llama_file * file, bool prefetch = true) {
+    llama_mmap(struct llama_file * file, bool prefetch = true, bool numa = false) {
+        (void) numa;
+
         size = file->size;
 
         HANDLE hFile = (HANDLE) _get_osfhandle(_fileno(file->fp));
@@ -244,8 +256,10 @@ struct llama_mmap {
 #else
     static constexpr bool SUPPORTED = false;
 
-    llama_mmap(struct llama_file *, bool prefetch = true) {
-        (void)prefetch;
+    llama_mmap(struct llama_file *, bool prefetch = true, bool numa = false) {
+        (void) prefetch;
+        (void) numa;
+
         throw std::runtime_error(std::string("mmap not supported"));
     }
 #endif