GC tunning simple patch ruby 1.9.2 p180

patch-1.9.2-gc.patch

diff --git a/gc.c b/gc.c
--- a/gc.c
+++ b/gc.c
@@ -77,6 +77,41 @@ void *alloca ();
 #ifndef GC_MALLOC_LIMIT
 #define GC_MALLOC_LIMIT 8000000
 #endif
+#define HEAP_MIN_SLOTS 10000
+#define FREE_MIN  4096
+
+static unsigned int initial_malloc_limit   = GC_MALLOC_LIMIT;
+static unsigned int initial_heap_min_slots = HEAP_MIN_SLOTS;
+static unsigned int initial_free_min       = FREE_MIN;
+
+static void set_gc_parameters()
+{
+   char *malloc_limit_ptr, *heap_min_slots_ptr, *free_min_ptr;
+
+   malloc_limit_ptr = getenv("RUBY_GC_MALLOC_LIMIT");
+   if ( malloc_limit_ptr != NULL ) {
+       int malloc_limit_i = atoi(malloc_limit_ptr);
+       if ( malloc_limit_i > 0 ) {
+	   initial_malloc_limit = malloc_limit_i;
+       }
+   }
+
+   heap_min_slots_ptr = getenv("RUBY_HEAP_MIN_SLOTS");
+   if ( heap_min_slots_ptr != NULL ) {
+       int heap_min_slots_i = atoi(heap_min_slots_ptr);
+       if ( heap_min_slots_i > 0 ) {
+	   initial_heap_min_slots = heap_min_slots_i;
+       }
+   }
+
+   free_min_ptr = getenv("RUBY_FREE_MIN");
+   if ( free_min_ptr != NULL ) {
+       int free_min_i = atoi(free_min_ptr);
+       if ( free_min_i > 0 ) {
+	   initial_free_min = free_min_i;
+       }
+   }
+}
 
 #define nomem_error GET_VM()->special_exceptions[ruby_error_nomemory]
 
@@ -283,9 +318,6 @@ struct heaps_slot {
     int finalize_flag;
 };
 
-#define HEAP_MIN_SLOTS 10000
-#define FREE_MIN  4096
-
 struct gc_list {
     VALUE *varptr;
     struct gc_list *next;
@@ -341,7 +373,7 @@ typedef struct rb_objspace {
 static int ruby_initial_gc_stress = 0;
 int *ruby_initial_gc_stress_ptr = &ruby_initial_gc_stress;
 #else
-static rb_objspace_t rb_objspace = {{GC_MALLOC_LIMIT}, {HEAP_MIN_SLOTS}};
+static rb_objspace_t rb_objspace = {{initial_malloc_limit}, {initial_heap_min_slots}};
 int *ruby_initial_gc_stress_ptr = &rb_objspace.gc_stress;
 #endif
 #define malloc_limit		objspace->malloc_params.limit
@@ -375,7 +407,7 @@ rb_objspace_alloc(void)
 {
     rb_objspace_t *objspace = malloc(sizeof(rb_objspace_t));
     memset(objspace, 0, sizeof(*objspace));
-    malloc_limit = GC_MALLOC_LIMIT;
+    malloc_limit = initial_malloc_limit;
     ruby_gc_stress = ruby_initial_gc_stress;
 
     return objspace;
@@ -958,7 +990,7 @@ init_heap(rb_objspace_t *objspace)
 {
     size_t add, i;
 
-    add = HEAP_MIN_SLOTS / HEAP_OBJ_LIMIT;
+    add = initial_heap_min_slots / HEAP_OBJ_LIMIT;
 
     if (!add) {
         add = 1;
@@ -1862,9 +1894,9 @@ gc_sweep(rb_objspace_t *objspace)
     do_heap_free = (size_t)((heaps_used * HEAP_OBJ_LIMIT) * 0.65);
     free_min = (size_t)((heaps_used * HEAP_OBJ_LIMIT)  * 0.2);
 
-    if (free_min < FREE_MIN) {
+    if (free_min < initial_free_min) {
 	do_heap_free = heaps_used * HEAP_OBJ_LIMIT;
-        free_min = FREE_MIN;
+        free_min = initial_free_min;
     }
 
     freelist = 0;
@@ -1926,7 +1958,7 @@ gc_sweep(rb_objspace_t *objspace)
     GC_PROF_SET_MALLOC_INFO;
     if (malloc_increase > malloc_limit) {
 	malloc_limit += (size_t)((malloc_increase - malloc_limit) * (double)live / (live + freed));
-	if (malloc_limit < GC_MALLOC_LIMIT) malloc_limit = GC_MALLOC_LIMIT;
+	if (malloc_limit < initial_malloc_limit) malloc_limit = initial_malloc_limit;
     }
     malloc_increase = 0;
     if (freed < free_min) {
@@ -2298,6 +2330,7 @@ Init_stack(volatile VALUE *addr)
 void
 Init_heap(void)
 {
+    set_gc_parameters();
     init_heap(&rb_objspace);
 }
 

There an issue about getters and setters method on GC module in ruby issue tracker: http://redmine.ruby-lang.org/issues/1047
It includes patch which provides GC.malloc_limit= and GC.heap_slots= http://redmine.ruby-lang.org/attachments/873/1047.patch

I add patch to the gist for methods GC.malloc_limit=, GC.heap_min_slots= and GC.free_min=.
Attention: patch is showed in the gist as first but should be applied after patch-1.9.2-gc.patch !

If I think right, then I don't need your second patch as long as I'm not familiar with the GC.stuff because these methods only additionals.

Is there a good "tutorial" about GC in Ruby? Do you have tips and tricks for manual optimization of Ruby's GC/memory stuff?

And: Maybe you should create a new gist with both patches bundled to one. So nobody has to think about right order ;o)

(If you do that, please link the new gist here. Thanks!)

Glad to see Matz has merged this into trunk ruby/ruby@eb807d4

I used this patch for 1.9.2-p180 and I did not get what I expected to have. Rails benchmarks doesn't show proper values. I used also trunk version of ruby and still nothing. Is it possible to receive some tips how to compile ruby 1.9.2 to have benchmarks in rails. I used 1.8.7 Enterprise Edition and results are ok. Thanks for any help.

I did not get what I expected to have

What did you expect? What did you get? How did you setup environment variables?

Well, this patch do not improve 1.9.2 as Enterprise Edition improve 1.8.7. This patch allows 1.9.2 to be not slower than REE for typical web application.
1.9.2 is faster than 1.8.7, but its garbage collector worser than REE's one.
With this patch (and properly set variables) simple web pages runs with 1.9.2 as fast as with REE, so that I can go with 1.9.2.

What is your experience?

Now I understand what this patch do.

I want to run performance tests for Rails and to have all information e.g. created objects, memory, how many times gc runs, I need GC patch. With patch for 1.8.7 I have this information. With this patch I only receive value for time.
Now I don't have access to my notes but I will write my environment settings.
Regarding my experience I've never look so close of ruby language and ruby implementation. I know what is GC and what this do.

Is it possible to receive from you some notes/comments about what is missing to receive benchmark results for Rails application?
If you want to look for ruby benchmarks please follow this link: http://guides.rubyonrails.org/performance_testing.html

Best Regards

This patch is not implement statistics.
I took as a basis that patch from railsbench repository.
You could analyse railsbench patch and create patch against 1.9.2 .
I did it only for GC parameters cause I were not interested in deep benchmark statistic. "Siege" is enough for me.

@asirad maybe you want to try out my full GC patches available here: https://github.com/skaes/rvm-patchsets

@skaes cool patch set :)

will this work on p290?

will what work on p290? My patchsets work https://github.com/skaes/rvm-patchsets

ah, didn't realize your patchset included this functionality. awesome, i'll check it out, thanks.