add a compile-time option to enable 4k page sizes (#52229)

We're suffering from heavy fragmentation in some of our workloads.

Add a build-time option to enable 4k pages (instead of 16k) in the GC,
since that improves memory utilization considerably for us.

Drawback is that this may increase the number of `madvise` system calls
in the sweeping phase by a factor of 4, but concurrent page sweeping
should help with some of that.

Author: d-netto

src/gc.h

@@ -33,8 +33,12 @@
 extern "C" {
 #endif
 
+#ifdef GC_SMALL_PAGE
+#define GC_PAGE_LG2 12 // log2(size of a page)
+#else
 #define GC_PAGE_LG2 14 // log2(size of a page)
-#define GC_PAGE_SZ (1 << GC_PAGE_LG2) // 16k
+#endif
+#define GC_PAGE_SZ (1 << GC_PAGE_LG2)
 #define GC_PAGE_OFFSET (JL_HEAP_ALIGNMENT - (sizeof(jl_taggedvalue_t) % JL_HEAP_ALIGNMENT))
 
 #define jl_malloc_tag ((void*)0xdeadaa01)
@@ -242,6 +246,23 @@ typedef struct {
     _Atomic(size_t) n_pages_allocd;
 } gc_fragmentation_stat_t;
 
+#ifdef GC_SMALL_PAGE
+#ifdef _P64
+#define REGION0_PG_COUNT (1 << 16)
+#define REGION1_PG_COUNT (1 << 18)
+#define REGION2_PG_COUNT (1 << 18)
+#define REGION0_INDEX(p) (((uintptr_t)(p) >> 12) & 0xFFFF) // shift by GC_PAGE_LG2
+#define REGION1_INDEX(p) (((uintptr_t)(p) >> 28) & 0x3FFFF)
+#define REGION_INDEX(p)  (((uintptr_t)(p) >> 46) & 0x3FFFF)
+#else
+#define REGION0_PG_COUNT (1 << 10)
+#define REGION1_PG_COUNT (1 << 10)
+#define REGION2_PG_COUNT (1 << 0)
+#define REGION0_INDEX(p) (((uintptr_t)(p) >> 12) & 0x3FF) // shift by GC_PAGE_LG2
+#define REGION1_INDEX(p) (((uintptr_t)(p) >> 22) & 0x3FF)
+#define REGION_INDEX(p)  (0)
+#endif
+#else
 #ifdef _P64
 #define REGION0_PG_COUNT (1 << 16)
 #define REGION1_PG_COUNT (1 << 16)
@@ -257,6 +278,7 @@ typedef struct {
 #define REGION1_INDEX(p) (((uintptr_t)(p) >> 22) & 0x3FF)
 #define REGION_INDEX(p)  (0)
 #endif
+#endif
 
 // define the representation of the levels of the page-table (0 to 2)
 typedef struct {

src/julia_internal.h

@@ -373,24 +373,48 @@ static const int jl_gc_sizeclasses[] = {
     144, 160, 176, 192, 208, 224, 240, 256,
 
     // the following tables are computed for maximum packing efficiency via the formula:
-    // pg = 2^14
+    // pg = GC_SMALL_PAGE ? 2^12 : 2^14
     // sz = (div.(pg-8, rng).÷16)*16; hcat(sz, (pg-8).÷sz, pg .- (pg-8).÷sz.*sz)'
 
+#ifdef GC_SMALL_PAGE
+    // rng = 15:-1:2 (14 pools)
+    272, 288, 304, 336, 368, 400, 448, 496, 576, 672, 816, 1008, 1360, 2032
+//  15, 14, 13, 12, 11, 10, 9, 8, 7, 6, 5, 4, 3, 2, /pool
+//  16, 64, 144, 64, 48, 96, 64, 128, 64, 64, 16, 64, 16, 32, bytes lost
+#else
     // rng = 60:-4:32 (8 pools)
     272, 288, 304, 336, 368, 400, 448, 496,
-//   60,  56,  53,  48,  44,  40,  36,  33, /pool
-//   64, 256, 272, 256, 192, 384, 256,  16, bytes lost
+//  60, 56, 53, 48, 44, 40, 36, 33, /pool
+//  64, 256, 272, 256, 192, 384, 256,  16, bytes lost
 
     // rng = 30:-2:16 (8 pools)
     544, 576, 624, 672, 736, 816, 896, 1008,
-//   30,  28,  26,  24,  22,  20,  18,  16, /pool
-//   64, 256, 160, 256, 192,  64, 256, 256, bytes lost
+//  30, 28, 26, 24, 22, 20, 18, 16, /pool
+//  64, 256, 160, 256, 192,  64, 256, 256, bytes lost
 
     // rng = 15:-1:8 (8 pools)
     1088, 1168, 1248, 1360, 1488, 1632, 1808, 2032
-//    15,   14,   13,   12,   11,   10,    9,    8, /pool
-//    64,   32,  160,   64,   16,   64,  112,  128, bytes lost
+//   15, 14, 13, 12, 11, 10, 9, 8, /pool
+//   64, 32, 160, 64, 16, 64, 112,  128, bytes lost
+#endif
 };
+#ifdef GC_SMALL_PAGE
+#ifdef _P64
+#  define JL_GC_N_POOLS 39
+#elif MAX_ALIGN == 8
+#  define JL_GC_N_POOLS 40
+#else
+#  define JL_GC_N_POOLS 41
+#endif
+#else
+#ifdef _P64
+#  define JL_GC_N_POOLS 49
+#elif MAX_ALIGN == 8
+#  define JL_GC_N_POOLS 50
+#else
+#  define JL_GC_N_POOLS 51
+#endif
+#endif
 static_assert(sizeof(jl_gc_sizeclasses) / sizeof(jl_gc_sizeclasses[0]) == JL_GC_N_POOLS, "");
 
 STATIC_INLINE int jl_gc_alignment(size_t sz) JL_NOTSAFEPOINT
@@ -417,7 +441,12 @@ JL_DLLEXPORT int jl_alignment(size_t sz) JL_NOTSAFEPOINT;
 
 // the following table is computed as:
 // [searchsortedfirst(jl_gc_sizeclasses, i) - 1 for i = 0:16:jl_gc_sizeclasses[end]]
-static const uint8_t szclass_table[] = {0, 1, 3, 5, 7, 9, 11, 13, 15, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 28, 29, 29, 30, 30, 31, 31, 31, 32, 32, 32, 33, 33, 33, 34, 34, 35, 35, 35, 36, 36, 36, 37, 37, 37, 37, 38, 38, 38, 38, 38, 39, 39, 39, 39, 39, 40, 40, 40, 40, 40, 40, 40, 41, 41, 41, 41, 41, 42, 42, 42, 42, 42, 43, 43, 43, 43, 43, 44, 44, 44, 44, 44, 44, 44, 45, 45, 45, 45, 45, 45, 45, 45, 46, 46, 46, 46, 46, 46, 46, 46, 46, 47, 47, 47, 47, 47, 47, 47, 47, 47, 47, 47, 48, 48, 48, 48, 48, 48, 48, 48, 48, 48, 48, 48, 48, 48};
+static const uint8_t szclass_table[] =
+#ifdef GC_SMALL_PAGE
+    {0,1,3,5,7,9,11,13,15,17,18,19,20,21,22,23,24,25,26,27,28,28,29,29,30,30,31,31,31,32,32,32,33,33,33,33,33,34,34,34,34,34,34,35,35,35,35,35,35,35,35,35,36,36,36,36,36,36,36,36,36,36,36,36,37,37,37,37,37,37,37,37,37,37,37,37,37,37,37,37,37,37,37,37,37,37,38,38,38,38,38,38,38,38,38,38,38,38,38,38,38,38,38,38,38,38,38,38,38,38,38,38,38,38,38,38,38,38,38,38,38,38,38,38,38,38,38,38};
+#else
+    {0,1,3,5,7,9,11,13,15,17,18,19,20,21,22,23,24,25,26,27,28,28,29,29,30,30,31,31,31,32,32,32,33,33,33,34,34,35,35,35,36,36,36,37,37,37,37,38,38,38,38,38,39,39,39,39,39,40,40,40,40,40,40,40,41,41,41,41,41,42,42,42,42,42,43,43,43,43,43,44,44,44,44,44,44,44,45,45,45,45,45,45,45,45,46,46,46,46,46,46,46,46,46,47,47,47,47,47,47,47,47,47,47,47,48,48,48,48,48,48,48,48,48,48,48,48,48,48};
+#endif
 static_assert(sizeof(szclass_table) == 128, "");
 
 STATIC_INLINE uint8_t JL_CONST_FUNC jl_gc_szclass(unsigned sz) JL_NOTSAFEPOINT

src/julia_threads.h

@@ -4,8 +4,8 @@
 #ifndef JL_THREADS_H
 #define JL_THREADS_H
 
-#include "work-stealing-queue.h"
 #include "julia_atomics.h"
+#include "work-stealing-queue.h"
 #ifndef _OS_WINDOWS_
 #include "pthread.h"
 #endif
@@ -161,14 +161,8 @@ typedef struct {
     arraylist_t *last_remset;
 
     // variables for allocating objects from pools
-#ifdef _P64
-#  define JL_GC_N_POOLS 49
-#elif MAX_ALIGN == 8
-#  define JL_GC_N_POOLS 50
-#else
-#  define JL_GC_N_POOLS 51
-#endif
-    jl_gc_pool_t norm_pools[JL_GC_N_POOLS];
+#define JL_GC_N_MAX_POOLS 51 // conservative. must be kept in sync with `src/julia_internal.h`
+    jl_gc_pool_t norm_pools[JL_GC_N_MAX_POOLS];
 
 #define JL_N_STACK_POOLS 16
     small_arraylist_t free_stacks[JL_N_STACK_POOLS];

src/options.h

@@ -78,6 +78,11 @@
 // OBJPROFILE counts objects by type
 // #define OBJPROFILE
 
+// pool allocator configuration options
+
+// GC_SMALL_PAGE allocates objects in 4k pages
+// #define GC_SMALL_PAGE
+
 
 // method dispatch profiling --------------------------------------------------