Bug #18871046 SET NUMA MEMPOLICY FOR OPTIMUM MYSQLD PERFORMANCE

Problem:

This is a contributed patch (bug#72811). The original contributor
states that, "Historically, the suggested way to run MySQL on NUMA
machines has been to run with "numactl --interleave=all" so that
things like the InnoDB buffer pool are spread across all NUMA
nodes." Currently InnoDB buffer pool is allocated without using
NUMA features.

Solution:

Provide an option (--innodb_numa_interleave) to allocate InnoDB buffer
pool using MPOL_INTERLEAVE memory policy of NUMA. What this patch does
is the following:

1.  If the option innodb_numa_interleave is enabled, then set the numa
    memory policy MPOL_INTERLEAVE for the mysqld process.
2.  Allocate the buffer pool.
3.  If the option innodb_numa_interleave is enabled, then set the numa
    memory policy back to MPOL_DEFAULT for the mysqld process.

So all allocations that happen between 1 and 3 will be using the
MPOL_INTERLEAVE memory policy.  This is a basic patch to make InnoDB
numa aware.  More fine grained control is possible and will be taken
up later.

rb#7309 approved by allen.

config.h.cmake

@@ -1,4 +1,4 @@
-/* Copyright (c) 2009, 2014, Oracle and/or its affiliates. All rights reserved.
+/* Copyright (c) 2009, 2015, Oracle and/or its affiliates. All rights reserved.
 
  This program is free software; you can redistribute it and/or modify
  it under the terms of the GNU General Public License as published by
@@ -664,5 +664,6 @@
 /* CPU information */
 
 #cmakedefine CPU_LEVEL1_DCACHE_LINESIZE @CPU_LEVEL1_DCACHE_LINESIZE@
+#cmakedefine HAVE_LIBNUMA 1
 
 #endif

configure.cmake

@@ -1,4 +1,4 @@
-# Copyright (c) 2009, 2014, Oracle and/or its affiliates. All rights reserved.
+# Copyright (c) 2009, 2015, Oracle and/or its affiliates. All rights reserved.
 # 
 # This program is free software; you can redistribute it and/or modify
 # it under the terms of the GNU General Public License as published by
@@ -1138,3 +1138,17 @@ SET(CMAKE_EXTRA_INCLUDE_FILES)
 CHECK_STRUCT_HAS_MEMBER("struct dirent" d_ino "dirent.h"  STRUCT_DIRENT_HAS_D_INO)
 CHECK_STRUCT_HAS_MEMBER("struct dirent" d_namlen "dirent.h"  STRUCT_DIRENT_HAS_D_NAMLEN)
 SET(SPRINTF_RETURNS_INT 1)
+
+CHECK_INCLUDE_FILES (numaif.h HAVE_NUMAIF_H)
+SET(WITH_NUMA 1 CACHE BOOL "Explicitly set NUMA memory allocation policy")
+IF(HAVE_NUMAIF_H AND WITH_NUMA)
+    SET(CMAKE_REQUIRED_LIBRARIES ${CMAKE_REQUIRED_LIBRARIES} numa)
+    CHECK_C_SOURCE_COMPILES(
+    "
+    #include <numaif.h>
+    int main()
+    {
+       set_mempolicy(MPOL_DEFAULT, 0, 0);
+    }"
+    HAVE_LIBNUMA)
+ENDIF()

mysql-test/include/have_numa.inc

@@ -0,0 +1,18 @@
+let $numa_support = `SELECT COUNT(VARIABLE_VALUE) = 1 FROM
+  INFORMATION_SCHEMA.GLOBAL_VARIABLES
+  WHERE VARIABLE_NAME='innodb_numa_interleave'`;
+
+if ( $numa_support == 0 )
+{
+    --skip Test requires: NUMA must be enabled
+}
+
+if ( $numa_support == 1 )
+{
+  let $numa = `SELECT @@GLOBAL.innodb_numa_interleave = 'OFF'`;
+  if ( $numa == 1 )
+  {
+    --skip Test requires: NUMA must be enabled
+  }
+}
+

mysql-test/suite/sys_vars/r/innodb_numa_interleave_basic.result

@@ -0,0 +1,10 @@
+SELECT @@GLOBAL.innodb_numa_interleave;
+@@GLOBAL.innodb_numa_interleave
+1
+SET @@GLOBAL.innodb_numa_interleave=off;
+ERROR HY000: Variable 'innodb_numa_interleave' is a read only variable
+SELECT @@GLOBAL.innodb_use_native_aio;
+@@GLOBAL.innodb_use_native_aio
+0
+SELECT @@SESSION.innodb_use_native_aio;
+ERROR HY000: Variable 'innodb_use_native_aio' is a GLOBAL variable

mysql-test/suite/sys_vars/t/innodb_numa_interleave_basic.test

@@ -0,0 +1,13 @@
+--source include/have_innodb.inc
+--source include/have_numa.inc
+
+SELECT @@GLOBAL.innodb_numa_interleave;
+
+--error ER_INCORRECT_GLOBAL_LOCAL_VAR
+SET @@GLOBAL.innodb_numa_interleave=off;
+
+SELECT @@GLOBAL.innodb_use_native_aio;
+
+--error ER_INCORRECT_GLOBAL_LOCAL_VAR
+SELECT @@SESSION.innodb_use_native_aio;
+

storage/innobase/CMakeLists.txt

@@ -1,4 +1,4 @@
-# Copyright (c) 2006, 2011, Oracle and/or its affiliates. All rights reserved.
+# Copyright (c) 2006, 2015, Oracle and/or its affiliates. All rights reserved.
 #
 # This program is free software; you can redistribute it and/or modify
 # it under the terms of the GNU General Public License as published by
@@ -29,6 +29,9 @@ IF(UNIX)
       ADD_DEFINITIONS(-DLINUX_NATIVE_AIO=1)
       LINK_LIBRARIES(aio)
     ENDIF()
+    IF(HAVE_LIBNUMA)
+      LINK_LIBRARIES(numa)
+    ENDIF()
   ELSEIF(CMAKE_SYSTEM_NAME MATCHES "HP*")
     ADD_DEFINITIONS("-DUNIV_HPUX")
   ELSEIF(CMAKE_SYSTEM_NAME STREQUAL "AIX")

storage/innobase/buf/buf0buf.cc

@@ -53,6 +53,10 @@ Created 11/5/1995 Heikki Tuuri
 #include "page0zip.h"
 #include "srv0mon.h"
 #include "buf0checksum.h"
+#ifdef HAVE_LIBNUMA
+#include <numa.h>
+#include <numaif.h>
+#endif // HAVE_LIBNUMA
 
 /*
 		IMPLEMENTATION OF THE BUFFER POOL
@@ -1442,6 +1446,21 @@ buf_pool_init(
 	ut_ad(n_instances <= MAX_BUFFER_POOLS);
 	ut_ad(n_instances == srv_buf_pool_instances);
 
+#ifdef HAVE_LIBNUMA
+	if (srv_numa_interleave) {
+		ib_logf(IB_LOG_LEVEL_INFO,
+			"Setting NUMA memory policy to MPOL_INTERLEAVE");
+		if (set_mempolicy(MPOL_INTERLEAVE,
+				  numa_all_nodes_ptr->maskp,
+				  numa_all_nodes_ptr->size) != 0) {
+			ib_logf(IB_LOG_LEVEL_WARN,
+				"Failed to set NUMA memory policy to"
+				" MPOL_INTERLEAVE (error: %s).",
+				strerror(errno));
+		}
+	}
+#endif // HAVE_LIBNUMA
+
 	buf_pool_ptr = (buf_pool_t*) mem_zalloc(
 		n_instances * sizeof *buf_pool_ptr);
 
@@ -1462,6 +1481,18 @@ buf_pool_init(
 
 	btr_search_sys_create(buf_pool_get_curr_size() / sizeof(void*) / 64);
 
+#ifdef HAVE_LIBNUMA
+	if (srv_numa_interleave) {
+		ib_logf(IB_LOG_LEVEL_INFO,
+			"Setting NUMA memory policy to MPOL_DEFAULT");
+		if (set_mempolicy(MPOL_DEFAULT, NULL, 0) != 0) {
+			ib_logf(IB_LOG_LEVEL_WARN,
+				"Failed to set NUMA memory policy to"
+				" MPOL_DEFAULT (error: %s).", strerror(errno));
+		}
+	}
+#endif // HAVE_LIBNUMA
+
 	return(DB_SUCCESS);
 }
 

storage/innobase/handler/ha_innodb.cc

@@ -16301,6 +16301,13 @@ static MYSQL_SYSVAR_BOOL(use_native_aio, srv_use_native_aio,
   "Use native AIO if supported on this platform.",
   NULL, NULL, TRUE);
 
+#ifdef HAVE_LIBNUMA
+static MYSQL_SYSVAR_BOOL(numa_interleave, srv_numa_interleave,
+  PLUGIN_VAR_NOCMDARG | PLUGIN_VAR_READONLY,
+  "Use NUMA interleave memory policy to allocate InnoDB buffer pool.",
+  NULL, NULL, FALSE);
+#endif // HAVE_LIBNUMA
+
 static MYSQL_SYSVAR_BOOL(api_enable_binlog, ib_binlog_enabled,
   PLUGIN_VAR_NOCMDARG | PLUGIN_VAR_READONLY,
   "Enable binlog for applications direct access InnoDB through InnoDB APIs",
@@ -16579,6 +16586,9 @@ static struct st_mysql_sys_var* innobase_system_variables[]= {
   MYSQL_SYSVAR(version),
   MYSQL_SYSVAR(use_sys_malloc),
   MYSQL_SYSVAR(use_native_aio),
+#ifdef HAVE_LIBNUMA
+  MYSQL_SYSVAR(numa_interleave),
+#endif // HAVE_LIBNUMA
   MYSQL_SYSVAR(change_buffering),
   MYSQL_SYSVAR(change_buffer_max_size),
 #if defined UNIV_DEBUG || defined UNIV_IBUF_DEBUG

storage/innobase/include/srv0srv.h

@@ -220,6 +220,7 @@ OS (provided we compiled Innobase with it in), otherwise we will
 use simulated aio we build below with threads.
 Currently we support native aio on windows and linux */
 extern my_bool	srv_use_native_aio;
+extern my_bool	srv_numa_interleave;
 #ifdef __WIN__
 extern ibool	srv_use_native_conditions;
 #endif /* __WIN__ */
@@ -878,6 +879,7 @@ struct srv_slot_t{
 #else /* !UNIV_HOTBACKUP */
 # define srv_use_adaptive_hash_indexes		FALSE
 # define srv_use_native_aio			FALSE
+# define srv_numa_interleave			FALSE
 # define srv_force_recovery			0UL
 # define srv_set_io_thread_op_info(t,info)	((void) 0)
 # define srv_reset_io_thread_op_info()		((void) 0)

storage/innobase/srv/srv0srv.cc

@@ -147,6 +147,7 @@ OS (provided we compiled Innobase with it in), otherwise we will
 use simulated aio we build below with threads.
 Currently we support native aio on windows and linux */
 UNIV_INTERN my_bool	srv_use_native_aio = TRUE;
+UNIV_INTERN my_bool	srv_numa_interleave = FALSE;
 
 #ifdef __WIN__
 /* Windows native condition variables. We use runtime loading / function