rts: Gradually return retained memory to the OS

Related to #19381 #19359 #14702

After a spike in memory usage we have been conservative about returning
allocated blocks to the OS in case we are still allocating a lot and would
end up just reallocating them. The result of this was that up to 4 * live_bytes
of blocks would be retained once they were allocated even if memory usage ended up
a lot lower.

For a heap of size ~1.5G, this would result in OS memory reporting 6G which is
both misleading and worrying for users.
In long-lived server applications this results in consistent high memory
usage when the live data size is much more reasonable (for example ghcide)

Therefore we have a new (2021) strategy which starts by retaining up to 4 * live_bytes
of blocks before gradually returning uneeded memory back to the OS on subsequent
major GCs which are NOT caused by a heap overflow.

Each major GC which is NOT caused by heap overflow increases the consec_idle_gcs
counter and the amount of memory which is retained is inversely proportional to this number.
By default the excess memory retained is
 oldGenFactor (controlled by -F) / 2 ^ (consec_idle_gcs * returnDecayFactor)

On a major GC caused by a heap overflow, the `consec_idle_gcs` variable is reset to 0
(as we could continue to allocate more, so retaining all the memory might make sense).

Therefore setting bigger values for `-Fd` makes the rate at which memory is returned slower.
Smaller values make it get returned faster. Setting `-Fd0` disables the
memory return completely, which is the behaviour of older GHC versions.

The default is `-Fd4` which results in the following scaling:

> mapM print [(x, 1/ (2**(x / 4))) | x <- [1 :: Double ..20]]
(1.0,0.8408964152537146)
(2.0,0.7071067811865475)
(3.0,0.5946035575013605)
(4.0,0.5)
(5.0,0.4204482076268573)
(6.0,0.35355339059327373)
(7.0,0.29730177875068026)
(8.0,0.25)
(9.0,0.21022410381342865)
(10.0,0.17677669529663687)
(11.0,0.14865088937534013)
(12.0,0.125)
(13.0,0.10511205190671433)
(14.0,8.838834764831843e-2)
(15.0,7.432544468767006e-2)
(16.0,6.25e-2)
(17.0,5.255602595335716e-2)
(18.0,4.4194173824159216e-2)
(19.0,3.716272234383503e-2)
(20.0,3.125e-2)

So after 13 consecutive GCs only 0.1 of the maximum memory used will be retained.

Further to this decay factor, the amount of memory we attempt to retain is
also influenced by the GC strategy for the oldest generation. If we are using
a copying strategy then we will need at least 2 * live_bytes for copying to take
place, so we always keep that much. If using compacting or nonmoving then we need a lower number,
so we just retain at least `1.2 * live_bytes` for some protection.

In future we might want to make this behaviour more aggressive, some
relevant literature is

> Ulan Degenbaev, Jochen Eisinger, Manfred Ernst, Ross McIlroy, and Hannes Payer. 2016. Idle time garbage collection scheduling. SIGPLAN Not. 51, 6 (June 2016), 570–583. DOI:https://doi.org/10.1145/2980983.2908106

which describes the "memory reducer" in the V8 javascript engine which
on an idle collection immediately returns as much memory as possible.

2 files changed, 75 insertions, 5 deletions

diff --git a/rts/sm/GC.c b/rts/sm/GC.c
index 2911aad7a0..1a71bd7bf0 100644
--- a/rts/sm/GC.c
+++ b/rts/sm/GC.c
@@ -120,6 +120,8 @@ bool unload_mark_needed;
  */
 static W_ g0_pcnt_kept = 30; // percentage of g0 live at last minor GC
 
+static int consec_idle_gcs = 0;
+
 /* Mut-list stats */
 #if defined(DEBUG)
 // For lack of a better option we protect mutlist_scav_stats with oldest_gen->sync
@@ -261,6 +263,7 @@ addMutListScavStats(const MutListScavStats *src,
 void
 GarbageCollect (uint32_t collect_gen,
                 const bool do_heap_census,
+                const bool is_overflow_gc,
                 const bool deadlock_detect,
                 uint32_t gc_type USED_IF_THREADS,
                 Capability *cap,
@@ -981,11 +984,26 @@ GarbageCollect (uint32_t collect_gen,
       }
 #endif
 
-      /* If the amount of data remains constant, next major GC we'll
-       * require (F+1)*live + prealloc. We leave (F+2)*live + prealloc
-       * in order to reduce repeated deallocation and reallocation. #14702
-       */
-      need = need_prealloc + (RtsFlags.GcFlags.oldGenFactor + 2) * need_live;
+      // Reset the counter if the major GC was caused by a heap overflow
+      consec_idle_gcs = is_overflow_gc ? 0 : consec_idle_gcs + 1;
+
+      // See Note [Scaling retained memory]
+      double scaled_factor =
+        RtsFlags.GcFlags.returnDecayFactor > 0
+          ? RtsFlags.GcFlags.oldGenFactor / pow(2, (float) consec_idle_gcs / RtsFlags.GcFlags.returnDecayFactor)
+          : RtsFlags.GcFlags.oldGenFactor;
+
+      debugTrace(DEBUG_gc, "factors: %f %d %f", RtsFlags.GcFlags.oldGenFactor, consec_idle_gcs, scaled_factor  );
+
+      // Unavoidable need depends on GC strategy
+      // * Copying need 2 * live
+      // * Compacting need 1.x * live (we choose 1.2)
+      // * Nonmoving needs ~ 1.x * live
+      double unavoidable_need_factor = (oldest_gen->compact || RtsFlags.GcFlags.useNonmoving)
+                                          ? 1.2 : 2;
+      W_ scaled_needed = (scaled_factor + unavoidable_need_factor) * need_live;
+      debugTrace(DEBUG_gc, "factors_2: %f %d", unavoidable_need_factor, scaled_needed);
+      need = need_prealloc + scaled_needed;
 
       /* Also, if user set heap size, do not drop below it.
        */
@@ -1003,6 +1021,7 @@ GarbageCollect (uint32_t collect_gen,
       need = BLOCKS_TO_MBLOCKS(need);
 
       got = mblocks_allocated;
+      debugTrace(DEBUG_gc,"Returning: %d %d", got, need);
 
       uint32_t returned = 0;
       if (got > need) {
@@ -2208,3 +2227,53 @@ bool doIdleGCWork(Capability *cap STG_UNUSED, bool all)
  * work_stealing is "mostly immutable". We set it to false when we begin the
  * final sequential collections, for the benefit of notifyTodoBlock.
  * */
+
+/* Note [Scaling retained memory]
+ * Tickets: #19381 #19359 #14702
+ *
+ * After a spike in memory usage we have been conservative about returning
+ * allocated blocks to the OS in case we are still allocating a lot and would
+ * end up just reallocating them. The result of this was that up to 4 * live_bytes
+ * of blocks would be retained once they were allocated even if memory usage ended up
+ * a lot lower.
+ *
+ * For a heap of size ~1.5G, this would result in OS memory reporting 6G which is
+ * both misleading and worrying for users.
+ * In long-lived server applications this results in consistent high memory
+ * usage when the live data size is much more reasonable (for example ghcide)
+ *
+ * Therefore we have a new (2021) strategy which starts by retaining up to 4 * live_bytes
+ * of blocks before gradually returning uneeded memory back to the OS on subsequent
+ * major GCs which are NOT caused by a heap overflow.
+ *
+ * Each major GC which is NOT caused by heap overflow increases the consec_idle_gcs
+ * counter and the amount of memory which is retained is inversely proportional to this number.
+ * By default the excess memory retained is
+ *  oldGenFactor (controlled by -F) / 2 ^ (consec_idle_gcs * returnDecayFactor)
+ *
+ * On a major GC caused by a heap overflow, the `consec_idle_gcs` variable is reset to 0
+ * (as we could continue to allocate more, so retaining all the memory might make sense).
+ *
+ * Therefore setting bigger values for `-Fd` makes the rate at which memory is returned slower.
+ * Smaller values make it get returned faster. Setting `-Fd0` means no additional memory
+ * is retained.
+ *
+ * The default is `-Fd4` which results in the following scaling:
+ *
+ * > mapM print [(x, 1/ (2**(x / 4))) | x <- [1 :: Double ..20]]
+ * (1.0,0.8408964152537146)
+ * ...
+ * (4.0,0.5)
+ * ...
+ * (12.0,0.125)
+ * ...
+ * (20.0,3.125e-2)
+ *
+ * So after 12 consecutive GCs only 0.1 of the maximum memory used will be retained.
+ *
+ * Further to this decay factor, the amount of memory we attempt to retain is
+ * also influenced by the GC strategy for the oldest generation. If we are using
+ * a copying strategy then we will need at least 2 * live_bytes for copying to take
+ * place, so we always keep that much. If using compacting or nonmoving then we need a lower number,
+ * so we just retain at least `1.2 * live_bytes` for some protection.
+ */
diff --git a/rts/sm/GC.h b/rts/sm/GC.h
index 239f281910..da90c61302 100644
--- a/rts/sm/GC.h
+++ b/rts/sm/GC.h
@@ -19,6 +19,7 @@
 
 void GarbageCollect (uint32_t collect_gen,
                      bool do_heap_census,
+                     bool is_overflow_gc,
                      bool deadlock_detect,
                      uint32_t gc_type,
                      Capability *cap,