ADBDEV-6156 Count startup memory of each process when using resource groups

src/backend/utils/mmgr/vmem_tracker.c

@@ -633,6 +633,12 @@ VmemTracker_ReleaseVmem(int64 toBeFreedRequested)
 	}
 }
 
+int32
+VmemTracker_GetStartupChunks(void)
+{
+	return startupChunks;
+}
+
 /*
  * Register the startup memory to vmem tracker.
  *
@@ -670,6 +676,8 @@ VmemTracker_RegisterStartupMemory(int64 bytes)
 	pg_atomic_add_fetch_u32((pg_atomic_uint32 *) segmentVmemChunks,
 							startupChunks);
 
+	ResGroupProcAddStartupChunks(startupChunks);
+
 	/*
 	 * Step 2, check if an OOM error should be raised by allocating 0 chunk.
 	 */
@@ -692,6 +700,8 @@ VmemTracker_UnregisterStartupMemory(void)
 		pg_atomic_sub_fetch_u32((pg_atomic_uint32 *) &MySessionState->sessionVmem,
 								startupChunks);
 
+	ResGroupProcSubStartupChunks(startupChunks);
+
 	trackedBytes -= startupBytes;
 	trackedVmemChunks -= startupChunks;
 

src/backend/utils/resgroup/resgroup.c

@@ -1594,6 +1594,30 @@ selfDetachResGroup(ResGroupData *group, ResGroupSlotData *slot)
 	selfUnsetGroup();
 }
 
+/* 
+ * Add startup memory before a resgroup is assigned. This memory
+ * will later be added to resgroup via selfAttachResGroup
+ */
+void
+ResGroupProcAddStartupChunks(int32 chunks)
+{
+	if (IsResGroupEnabled())
+		self->memUsage += chunks;
+}
+
+/* 
+ * Sub startup memory at cleanup. This memory should already been
+ * subtracted from a resource group via selfDetachResGroup.
+ * Actually, this is not needed because a running process will always have
+ * startup memory consumpion, but let it be just for symmetry.
+ */
+void
+ResGroupProcSubStartupChunks(int32 chunks)
+{
+	if (IsResGroupEnabled())
+		self->memUsage -= chunks;
+}
+
 /*
  * Initialize the members of a slot
  */

src/include/utils/resgroup.h

@@ -235,6 +235,9 @@ extern Oid ResGroupGetGroupIdBySessionId(int sessionId);
 extern char *getCpuSetByRole(const char *cpuset);
 extern void checkCpuSetByRole(const char *cpuset);
 
+extern void ResGroupProcAddStartupChunks(int32 chunks);
+extern void ResGroupProcSubStartupChunks(int32 chunks);
+
 #define LOG_RESGROUP_DEBUG(...) \
 	do {if (Debug_resource_group) elog(__VA_ARGS__); } while(false);
 

src/include/utils/vmem_tracker.h

@@ -66,6 +66,7 @@ extern void VmemTracker_UnregisterStartupMemory(void);
 extern void VmemTracker_RequestWaiver(int64 waiver_bytes);
 extern void VmemTracker_ResetWaiver(void);
 extern int64 VmemTracker_Fault(int32 reason, int64 arg);
+extern int32 VmemTracker_GetStartupChunks(void);
 
 extern int32 RedZoneHandler_GetRedZoneLimitChunks(void);
 extern int32 RedZoneHandler_GetRedZoneLimitMB(void);

src/test/isolation2/input/resgroup/resgroup_move_query.source

@@ -11,6 +11,10 @@
 --
 -- end_matchsubs
 
+-- start_ignore
+! gpstop -rai;
+-- end_ignore
+
 CREATE OR REPLACE FUNCTION resGroupPalloc(float) RETURNS int AS
 '@abs_builddir@/../regress/regress@DLSUFFIX@', 'resGroupPalloc'
 LANGUAGE C READS SQL DATA;
@@ -135,7 +139,7 @@ SELECT num_running FROM gp_toolkit.gp_resgroup_status WHERE rsgname='rg_move_que
 1&: SELECT pg_sleep(3);
 2: SET ROLE role_move_query_mem_small;
 2: BEGIN;
-2: SELECT hold_memory_by_percent_on_qe(1,0.1);
+2: SELECT hold_memory_by_percent_on_qe(1,0.2);
 3&: SELECT gp_toolkit.pg_resgroup_move_query(pid, 'rg_move_query') FROM pg_stat_activity WHERE query LIKE '%hold_memory_by_percent_on_qe%' AND rsgname='rg_move_query_mem_small';
 1<:
 -- connection 1 finished, it will wake up connection 3
@@ -150,7 +154,7 @@ SELECT num_running FROM gp_toolkit.gp_resgroup_status WHERE rsgname='rg_move_que
 1: ALTER RESOURCE GROUP rg_move_query SET memory_limit 0;
 1: SET ROLE role_move_query_mem_small;
 1: BEGIN;
-1: SELECT hold_memory_by_percent_on_qe(1,0.1);
+1: SELECT hold_memory_by_percent_on_qe(1,0.2);
 2: SELECT gp_toolkit.pg_resgroup_move_query(pid, 'rg_move_query') FROM pg_stat_activity WHERE query LIKE '%hold_memory_by_percent_on_qe%' AND rsgname='rg_move_query_mem_small';
 2: SELECT is_session_in_group(pid, 'rg_move_query') FROM pg_stat_activity WHERE query LIKE '%hold_memory_by_percent_on_qe%' AND state = 'idle in transaction';
 1: END;

src/test/isolation2/input/resgroup/resgroup_startup_memory.source

@@ -0,0 +1,30 @@
+-- start_ignore
+! gpconfig -c runaway_detector_activation_percent -v 20;
+! gpstop -rai;
+-- end_ignore
+
+CREATE OR REPLACE FUNCTION resGroupPalloc(float) RETURNS int AS
+'@abs_builddir@/../regress/regress@DLSUFFIX@', 'resGroupPalloc'
+LANGUAGE C READS SQL DATA;
+
+CREATE OR REPLACE FUNCTION resGroupPallocIgnoreStartup(float) RETURNS int AS
+'@abs_builddir@/../regress/regress@DLSUFFIX@', 'resGroupPallocIgnoreStartup'
+LANGUAGE C READS SQL DATA;
+
+ALTER RESOURCE GROUP default_group SET memory_limit 20;
+ALTER RESOURCE GROUP admin_group SET memory_limit 10;
+ALTER RESOURCE GROUP admin_group set concurrency 1;
+
+5: select * from gp_toolkit.gp_resgroup_config;
+5q:
+
+-- This query will be killed by redzone because resGroupPallocIgnoreStartup just allocates
+-- the memory size of all resgroup memory. Before startup chunks were considered, a group had no memory tracked
+-- at the moment we called resGroupPalloc, so this case shows that now a group tracks startup memory.
+5: SELECT resGroupPallocIgnoreStartup(2.39) FROM gp_dist_random('gp_id') WHERE gp_segment_id = 1;
+5q:
+
+-- This won't fail because now resGroupPalloc subtracts startup chunks from the requested amount
+-- of memory to keep all previews tests intact. Leaving resGroupPalloc without changes may break them.
+5: SELECT resGroupPalloc(2.39) FROM gp_dist_random('gp_id') WHERE gp_segment_id = 1;
+5q:

src/test/isolation2/isolation2_resgroup_schedule

@@ -59,4 +59,6 @@ test: resgroup/resgroup_dumpinfo
 # test larget group id
 test: resgroup/resgroup_large_group_id
 
+test: resgroup/resgroup_startup_memory
+
 test: resgroup/disable_resgroup

src/test/isolation2/output/resgroup/resgroup_bypass.source

@@ -161,7 +161,7 @@ SELECT * FROM memory_result;
  rsgname        | ismaster | avg_mem 
 ----------------+----------+---------
  rg_bypass_test | 0        | 0.0     
- rg_bypass_test | 1        | 12.0    
+ rg_bypass_test | 1        | 24.0    
 (2 rows)
 61: SELECT * FROM eat_memory_on_qd_large;
 ERROR:  Out of memory
@@ -202,8 +202,8 @@ BEGIN
 SELECT * FROM memory_result;
  rsgname        | ismaster | avg_mem 
 ----------------+----------+---------
- rg_bypass_test | 0        | 4.0     
- rg_bypass_test | 1        | 0.0     
+ rg_bypass_test | 0        | 16.0    
+ rg_bypass_test | 1        | 12.0    
 (2 rows)
 61: SELECT * FROM eat_memory_on_one_slice;
  count 
@@ -213,8 +213,8 @@ SELECT * FROM memory_result;
 SELECT * FROM memory_result;
  rsgname        | ismaster | avg_mem 
 ----------------+----------+---------
- rg_bypass_test | 0        | 8.0     
- rg_bypass_test | 1        | 0.0     
+ rg_bypass_test | 0        | 20.0    
+ rg_bypass_test | 1        | 12.0    
 (2 rows)
 61: SELECT * FROM eat_memory_on_one_slice;
 ERROR:  Out of memory  (seg0 slice1 127.0.0.1:25432 pid=336)
@@ -234,7 +234,7 @@ SELECT * FROM memory_result;
  rsgname        | ismaster | avg_mem 
 ----------------+----------+---------
  rg_bypass_test | 0        | 0.0     
- rg_bypass_test | 1        | 0.0     
+ rg_bypass_test | 1        | 12.0    
 (2 rows)
 61q: ... <quitting>
 
@@ -256,8 +256,8 @@ BEGIN
 SELECT * FROM memory_result;
  rsgname        | ismaster | avg_mem 
 ----------------+----------+---------
- rg_bypass_test | 0        | 4.0     
- rg_bypass_test | 1        | 0.0     
+ rg_bypass_test | 0        | 16.0    
+ rg_bypass_test | 1        | 12.0    
 (2 rows)
 61: SELECT * FROM eat_memory_on_slices;
  count 
@@ -267,8 +267,8 @@ SELECT * FROM memory_result;
 SELECT * FROM memory_result;
  rsgname        | ismaster | avg_mem 
 ----------------+----------+---------
- rg_bypass_test | 0        | 8.0     
- rg_bypass_test | 1        | 0.0     
+ rg_bypass_test | 0        | 20.0    
+ rg_bypass_test | 1        | 12.0    
 (2 rows)
 61: SELECT * FROM eat_memory_on_slices;
 ERROR:  Out of memory  (seg0 slice2 127.0.0.1:25432 pid=354)
@@ -288,7 +288,7 @@ SELECT * FROM memory_result;
  rsgname        | ismaster | avg_mem 
 ----------------+----------+---------
  rg_bypass_test | 0        | 0.0     
- rg_bypass_test | 1        | 0.0     
+ rg_bypass_test | 1        | 12.0    
 (2 rows)
 61q: ... <quitting>
 

src/test/isolation2/output/resgroup/resgroup_memory_statistic.source

@@ -80,7 +80,7 @@ SELECT * FROM memory_result;
  rsgname         | ismaster | avg_mem 
 -----------------+----------+---------
  rg1_memory_test | 0        | 0.0     
- rg1_memory_test | 1        | 20.0    
+ rg1_memory_test | 1        | 30.0    
  rg2_memory_test | 0        | 0.0     
  rg2_memory_test | 1        | 0.0     
 (4 rows)
@@ -144,8 +144,8 @@ BEGIN
 SELECT * FROM memory_result;
  rsgname         | ismaster | avg_mem 
 -----------------+----------+---------
- rg1_memory_test | 0        | 20.0    
- rg1_memory_test | 1        | 0.0     
+ rg1_memory_test | 0        | 30.0    
+ rg1_memory_test | 1        | 10.0    
  rg2_memory_test | 0        | 0.0     
  rg2_memory_test | 1        | 0.0     
 (4 rows)

src/test/isolation2/output/resgroup/resgroup_move_query.source

@@ -11,6 +11,10 @@
 --
 -- end_matchsubs
 
+-- start_ignore
+! gpstop -rai;
+-- end_ignore
+
 CREATE OR REPLACE FUNCTION resGroupPalloc(float) RETURNS int AS '@abs_builddir@/../regress/regress@DLSUFFIX@', 'resGroupPalloc' LANGUAGE C READS SQL DATA;
 CREATE
 
@@ -185,7 +189,7 @@ SET
 SET
 2: BEGIN;
 BEGIN
-2: SELECT hold_memory_by_percent_on_qe(1,0.1);
+2: SELECT hold_memory_by_percent_on_qe(1,0.2);
  hold_memory_by_percent_on_qe 
 ------------------------------
  0                            
@@ -220,7 +224,7 @@ ALTER
 SET
 1: BEGIN;
 BEGIN
-1: SELECT hold_memory_by_percent_on_qe(1,0.1);
+1: SELECT hold_memory_by_percent_on_qe(1,0.2);
  hold_memory_by_percent_on_qe 
 ------------------------------
  0                            

src/test/isolation2/output/resgroup/resgroup_startup_memory.source

@@ -0,0 +1,41 @@
+-- start_ignore
+! gpconfig -c runaway_detector_activation_percent -v 20;
+! gpstop -rai;
+-- end_ignore
+
+CREATE OR REPLACE FUNCTION resGroupPalloc(float) RETURNS int AS '@abs_builddir@/../regress/regress@DLSUFFIX@', 'resGroupPalloc' LANGUAGE C READS SQL DATA;
+CREATE
+
+CREATE OR REPLACE FUNCTION resGroupPallocIgnoreStartup(float) RETURNS int AS '@abs_builddir@/../regress/regress@DLSUFFIX@', 'resGroupPallocIgnoreStartup' LANGUAGE C READS SQL DATA;
+CREATE
+
+ALTER RESOURCE GROUP default_group SET memory_limit 20;
+ALTER
+ALTER RESOURCE GROUP admin_group SET memory_limit 10;
+ALTER
+ALTER RESOURCE GROUP admin_group set concurrency 1;
+ALTER
+
+5: select * from gp_toolkit.gp_resgroup_config;
+ groupid | groupname     | concurrency | cpu_rate_limit | memory_limit | memory_shared_quota | memory_spill_ratio | memory_auditor | cpuset 
+---------+---------------+-------------+----------------+--------------+---------------------+--------------------+----------------+--------
+ 6437    | default_group | 20          | 30             | 20           | 80                  | 10                 | vmtracker      | -1     
+ 6438    | admin_group   | 1           | 10             | 10           | 80                  | 10                 | vmtracker      | -1     
+(2 rows)
+5q: ... <quitting>
+
+-- This query will be killed by redzone because resGroupPallocIgnoreStartup just allocates
+-- the memory size of all resgroup memory. Before startup chunks were considered, a group had no memory tracked
+-- at the moment we called resGroupPalloc, so this case shows that now a group tracks startup memory.
+5: SELECT resGroupPallocIgnoreStartup(2.39) FROM gp_dist_random('gp_id') WHERE gp_segment_id = 1;
+ERROR:  Canceling query because of high VMEM usage. current group id is 6438, group memory usage 175 MB, group shared memory quota is 55 MB, slot memory quota is 13 MB, global freechunks memory is 371 MB, global safe memory threshold is 382 MB (runaway_cleaner.c:197)  (seg1 slice1 10.92.43.77:6003 pid=29522) (runaway_cleaner.c:197)
+5q: ... <quitting>
+
+-- This won't fail because now resGroupPalloc subtracts startup chunks from the requested amount
+-- of memory to keep all previews tests intact. Leaving resGroupPalloc without changes may break them.
+5: SELECT resGroupPalloc(2.39) FROM gp_dist_random('gp_id') WHERE gp_segment_id = 1;
+ resgrouppalloc 
+----------------
+ 0              
+(1 row)
+5q: ... <quitting>

src/test/isolation2/sql/.gitignore

@@ -10,3 +10,4 @@ resgroup_bypass.sql
 resgroup_cpuset.sql
 gp_collation.sql
 distributed_snapshot.sql
+resgroup_startup_memory.sql