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 <[email protected]>

examples/common.cpp

@@ -343,6 +343,8 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
             params.use_mmap = false;
         } else if (arg == "--mtest") {
             params.mem_test = true;
+        } else if (arg == "--numa") {
+            params.numa = true;
         } else if (arg == "--export") {
             params.export_cgraph = true;
         } else if (arg == "--verbose-prompt") {
@@ -488,6 +490,9 @@ void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) {
     if (llama_mmap_supported()) {
         fprintf(stderr, "  --no-mmap             do not memory-map model (slower load but may reduce pageouts if not using mlock)\n");
     }
+    fprintf(stderr, "  --numa                attempt optimizations that help on some NUMA systems\n");
+    fprintf(stderr, "                        if run without this previously, it is recommended to drop the system page cache before using this\n");
+    fprintf(stderr, "                        see https://github.com/ggerganov/llama.cpp/issues/1437\n");
 #ifdef LLAMA_SUPPORTS_GPU_OFFLOAD
     fprintf(stderr, "  -ngl N, --n-gpu-layers N\n");
     fprintf(stderr, "                        number of layers to store in VRAM\n");

examples/common.h

@@ -76,6 +76,7 @@ struct gpt_params {
     bool use_mmap          = true;  // use mmap for faster loads
     bool use_mlock         = false; // use mlock to keep model in memory
     bool mem_test          = false; // compute maximum memory usage
+    bool numa              = false; // attempt optimizations that help on some NUMA systems
     bool export_cgraph     = false; // export the computation graph
     bool verbose_prompt    = false; // print prompt tokens before generation
 };

examples/embedding/embedding.cpp

@@ -35,7 +35,7 @@ int main(int argc, char ** argv) {
         params.prompt = gpt_random_prompt(rng);
     }
 
-    llama_init_backend();
+    llama_init_backend(params.numa);
 
     llama_model * model;
     llama_context * ctx;

examples/main/README.md

@@ -262,6 +262,10 @@ These options help improve the performance and memory usage of the LLaMA models.
 
 -   `--no-mmap`: Do not memory-map the model. By default, models are mapped into memory, which allows the system to load only the necessary parts of the model as needed. However, if the model is larger than your total amount of RAM or if your system is low on available memory, using mmap might increase the risk of pageouts, negatively impacting performance. Disabling mmap results in slower load times but may reduce pageouts if you're not using `--mlock`. Note that if the model is larger than the total amount of RAM, turning off mmap would prevent the model from loading at all.
 
+### NUMA support
+
+-   `--numa`: Attempt optimizations that help on some systems with non-uniform memory access. This currently consists of pinning an equal proportion of the threads to the cores on each NUMA node, and disabling prefetch and readahead for mmap. The latter causes mapped pages to be faulted in on first access instead of all at once, and in combination with pinning threads to NUMA nodes, more of the pages end up on the NUMA node where they are used. Note that if the model is already in the system page cache, for example because of a previous run without this option, this will have little effect unless you drop the page cache first. This can be done by rebooting the system or on Linux by writing '3' to '/proc/sys/vm/drop\_caches' as root.
+
 ### Memory Float 32
 
 -   `--memory-f32`: Use 32-bit floats instead of 16-bit floats for memory key+value. This doubles the context memory requirement and cached prompt file size but does not appear to increase generation quality in a measurable way. Not recommended.

examples/main/main.cpp

@@ -105,7 +105,7 @@ int main(int argc, char ** argv) {
         params.prompt = gpt_random_prompt(rng);
     }
 
-    llama_init_backend();
+    llama_init_backend(params.numa);
 
     llama_model * model;
     llama_context * ctx;

examples/perplexity/perplexity.cpp

@@ -147,7 +147,7 @@ int main(int argc, char ** argv) {
         params.prompt = gpt_random_prompt(rng);
     }
 
-    llama_init_backend();
+    llama_init_backend(params.numa);
 
     llama_model * model;
     llama_context * ctx;

examples/quantize/quantize.cpp

@@ -180,7 +180,7 @@ int main(int argc, char ** argv) {
         usage(argv[0]);
     }
 
-    llama_init_backend();
+    llama_init_backend(false);
 
     // parse command line arguments
     const std::string fname_inp = argv[arg_idx];

examples/server/server.cpp

@@ -789,7 +789,7 @@ int main(int argc, char ** argv) {
         params.model_alias = params.model;
     }
 
-    llama_init_backend();
+    llama_init_backend(params.numa);
 
     LOG_INFO("build info", {
         { "build", BUILD_NUMBER },

examples/simple/simple.cpp

@@ -66,7 +66,7 @@ int main(int argc, char ** argv)
     // Init LLM :
     //---------------------------------
 
-    llama_init_backend();
+    llama_init_backend(params.numa);
 
     llama_model * model;
     llama_context * ctx;