Skip to contentSkip to conversation

sysv ipc: scale msgmnb with the number of cpus

RFC In review [email protected] 3 participants 11 messages Jun 6, 2008 – Jun 10, 2008 Documentation

lore
Share & tools
lore
#0 lore raw

cover sysv ipc: scale msgmnb with the number of cpus

The size in bytes of a SysV IPC message queue, msgmnb, is too small for large machines, but we don't want to bloat small machines

Several methods are used already to modify (mainly increase) msgmnb:
. distribution specific patch
. system wide sysctl.conf
. application specific tuning via /proc/sys/kernel/msgmnb

Integrating this series would:
. reflect hardware and software evolutions and diversity, . reduce configuration/tuning for the applications.

Here is the timeline of the evolution of MSG* #defines:
Year 1994 1999 1999 2008
Version 1.0 2.3.27 2.3.30 2.6.24
#define MSGMNI 128 128 16 16
#define MSGMAX 4056 8192 8192 8192
#define MSGMNB 16384 16384 16384 16384

This patch series scales msgmnb, with respect to the number of cpus/cores for larger machines. For uniprocessor machines the value does not increase.

This series is similar to (and depends on) the series which scales msgmni, the number of IPC message queue identifiers, to the amount of low memory.
While Nadia's previous series scaled msgmni along the memory axis, hence the message pool (msgmni x msgmnb), this series uses a second
axis: the number of online CPUs.
As well as covering the (cpu,memory) space of machines size, this reflects the parallelism allowed by lockless send/receive for in-flight messages in queues (msgmnb / msgmax messages).

The initial scaling is done at initialization of the ipc namespace. Furthermore, the value becomes dynamic with respect to cpu hotplug.

The msgmni and msgmnb values become dependent, as the value of msgmni is computed with respect to the value of msgmnb.

The series is as follows:
. patch 1 introduces the scaling function
. patch 2 deals with cpu hotplug
. patch 3 allows user space to disable the scaling mechanism . patch 4 allows user space to reenable the scaling mechanism . patch 5 finer grain disabling/reenabling scaling mechanism
(disconnect msgmnb and msgmni)
. patch 6 adds documentation

---

The series applies to 2.6.26-rc2-mm1 + patch suppressing KERN_INFO messages as discussed at:
http://article.gmane.org/gmane.linux.kernel/686229
"[PATCH 1/1] Only output msgmni value at boot time"
(in mmotm: ipc-only-output-msgmni-value-at-boot-time.patch)

The plan would be to have this ready for the 2.6.27 merge window if there are no objections.

 Documentation/sysctl/kernel.txt |   27 ++++++++++++++++++++++
 include/linux/ipc_namespace.h   |    4 ++-
 include/linux/msg.h             |    5 ++++
 ipc/ipc_sysctl.c                |   48 ++++++++++++++++++++++++++++++----------
 ipc/ipcns_notifier.c            |   23 +++++++------------
 ipc/msg.c                       |   25 +++++++++++++++++---
 ipc/util.c                      |   28 +++++++++++++++++++++++
 ipc/util.h                      |    1
 8 files changed, 131 insertions(+), 30 deletions(-)
Signature · 2 lines
Solofo Ramangalahy
Bull SA.
#1 lore raw

RFC 1/6 sysv ipc: scale msgmnb to the number of cpus

From: Solofo Ramangalahy <[email protected]>

Initialize msgmnb value to
min(MSGMNB * num_online_cpus(), MSGMNB * MSG_CPU_SCALE) to increase the default value for larger machines.

MSG_CPU_SCALE scaling factor is defined to be 4, as 16384 x 4 = 65536 is an already used and recommended value.

The msgmni value is made dependant of msgmnb to keep the memory dedicated to message queues within the 1/MSG_MEM_SCALE of lowmem bound.

Unlike msgmni, the value is not scaled (down) with respect to the number of ipc namespaces for simplicity.

---
 include/linux/msg.h |    5 +++++
 ipc/msg.c           |   19 +++++++++++++++----
 2 files changed, 20 insertions(+), 4 deletions(-)
ipc/msg.c +15 -4
Index: b/ipc/msg.c
===================================================================
@@ -38,6 +38,7 @@
 #include <linux/rwsem.h>
 #include <linux/nsproxy.h>
 #include <linux/ipc_namespace.h>
+#include <linux/cpumask.h>
 
 #include <asm/current.h>
 #include <asm/uaccess.h>
@@ -92,7 +93,7 @@ void recompute_msgmni(struct ipc_namespa
 
 	si_meminfo(&i);
 	allowed = (((i.totalram - i.totalhigh) / MSG_MEM_SCALE) * i.mem_unit)
-		/ MSGMNB;
+		/ ns->msg_ctlmnb;
 	nb_ns = atomic_read(&nr_ipc_ns);
 	allowed /= nb_ns;
 
@@ -108,11 +109,21 @@ void recompute_msgmni(struct ipc_namespa
 
 	ns->msg_ctlmni = allowed;
 }
+/*
+ * Scale msgmnb with the number of online cpus, up to 4x MSGMNB.
+ * ns->msg_ctlmnb cannot be assigned zero because of division in
+ * recompute_msgmni()
+ */
+void recompute_msgmnb(struct ipc_namespace *ns)
+{
+	ns->msg_ctlmnb =
+		min(MSGMNB * num_online_cpus(), MSGMNB * MSG_CPU_SCALE);
+}
 
 void msg_init_ns(struct ipc_namespace *ns)
 {
 	ns->msg_ctlmax = MSGMAX;
-	ns->msg_ctlmnb = MSGMNB;
+	recompute_msgmnb(ns);
 
 	recompute_msgmni(ns);
 
@@ -132,8 +143,8 @@ void __init msg_init(void)
 {
 	msg_init_ns(&init_ipc_ns);
 
-	printk(KERN_INFO "msgmni has been set to %d\n",
-		init_ipc_ns.msg_ctlmni);
+	printk(KERN_INFO "msgmni has been set to %d, msgmnb to %d\n",
+	       init_ipc_ns.msg_ctlmni, init_ipc_ns.msg_ctlmnb);
 
 	ipc_init_proc_interface("sysvipc/msg",
 				"       key      msqid perms      cbytes       qnum lspid lrpid   uid   gid  cuid  cgid      stime      rtime      ctime\n",
include/linux/msg.h +5 -0
Index: b/include/linux/msg.h
===================================================================
@@ -58,6 +58,11 @@ struct msginfo {
  * more than 16 GB : msgmni = 32K (IPCMNI)
  */
 #define MSG_MEM_SCALE 32
+/*
+ * Scaling factor to compute msgmnb:
+ * ns->msg_ctlmnb is between MSGMNB and MSGMNB * MSG_CPU_SCALE
+ */
+#define MSG_CPU_SCALE 4
 
 #define MSGMNI    16   /* <= IPCMNI */     /* max # of msg queue identifiers */
 #define MSGMAX  8192   /* <= INT_MAX */   /* max size of message (bytes) */
Signature · 2 lines
Solofo Ramangalahy
Bull SA.
#2 lore raw

RFC 2/6 sysv ipc: recompute msgmnb (and msgmni) on cpu hotplug addition and removal

In reply to [email protected]

From: Solofo Ramangalahy <[email protected]>

As msgmnb is scaled wrt. online cpus, cpu hotplug events should grow and shrink the value.

Like msgmni with ipc_memory_callback(), the ipc_cpu_callback() function triggers msgmnb recomputation.

---
 include/linux/ipc_namespace.h |    1 +
 ipc/ipcns_notifier.c          |   14 +++++++++-----
 ipc/util.c                    |   28 ++++++++++++++++++++++++++++
 ipc/util.h                    |    1 +
 4 files changed, 39 insertions(+), 5 deletions(-)
include/linux/ipc_namespace.h +1 -0
Index: b/include/linux/ipc_namespace.h
===================================================================
@@ -12,6 +12,7 @@
 #define IPCNS_MEMCHANGED   0x00000001   /* Notify lowmem size changed */
 #define IPCNS_CREATED  0x00000002   /* Notify new ipc namespace created */
 #define IPCNS_REMOVED  0x00000003   /* Notify ipc namespace removed */
+#define IPCNS_CPUCHANGED  0x00000004   /* Notify cpu hotplug addition/removal */
 
 #define IPCNS_CALLBACK_PRI 0
 
ipc/util.c +28 -0
Index: b/ipc/util.c
===================================================================
@@ -34,6 +34,7 @@
 #include <linux/nsproxy.h>
 #include <linux/rwsem.h>
 #include <linux/memory.h>
+#include <linux/cpu.h>
 #include <linux/ipc_namespace.h>
 
 #include <asm/unistd.h>
@@ -96,6 +97,32 @@ static int ipc_memory_callback(struct no
 
 #endif /* CONFIG_MEMORY_HOTPLUG */
 
+#ifdef CONFIG_HOTPLUG_CPU
+
+static void ipc_cpu_notifier(struct work_struct *work)
+{
+	ipcns_notify(IPCNS_CPUCHANGED);
+}
+
+static DECLARE_WORK(ipc_cpu_wq, ipc_cpu_notifier);
+
+static int __cpuinit ipc_cpu_callback(struct notifier_block *nfb,
+				    unsigned long action, void *hcpu)
+{
+	switch (action) {
+	case CPU_ONLINE:
+	case CPU_ONLINE_FROZEN:
+	case CPU_DEAD:
+	case CPU_DEAD_FROZEN:
+		schedule_work(&ipc_cpu_wq);
+		break;
+	default:
+		break;
+	}
+	return NOTIFY_OK;
+}
+
+#endif /* CONFIG_HOTPLUG_CPU */
 /**
  *	ipc_init	-	initialise IPC subsystem
  *
@@ -112,6 +139,7 @@ static int __init ipc_init(void)
 	msg_init();
 	shm_init();
 	hotplug_memory_notifier(ipc_memory_callback, IPC_CALLBACK_PRI);
+	hotcpu_notifier(ipc_cpu_callback, IPC_CALLBACK_PRI);
 	register_ipcns_notifier(&init_ipc_ns);
 	return 0;
 }
ipc/ipcns_notifier.c +9 -5
Index: b/ipc/ipcns_notifier.c
===================================================================
@@ -26,16 +26,20 @@ static int ipcns_callback(struct notifie
 				unsigned long action, void *arg)
 {
 	struct ipc_namespace *ns;
-
+	ns = container_of(self, struct ipc_namespace, ipcns_nb);
 	switch (action) {
+	case IPCNS_CPUCHANGED:
+		/*
+		 * Fall through.
+		 * We do not scale msgmnb with IPC namespace
+		 * add/remove for simplicity (adjustment of the
+		 * message pool is done indirectly via msgmni).
+		 */
+		recompute_msgmnb(ns);
 	case IPCNS_MEMCHANGED:   /* amount of lowmem has changed */
 	case IPCNS_CREATED:
 	case IPCNS_REMOVED:
 		/*
-		 * It's time to recompute msgmni
-		 */
-		ns = container_of(self, struct ipc_namespace, ipcns_nb);
-		/*
 		 * No need to get a reference on the ns: the 1st job of
 		 * free_ipc_ns() is to unregister the callback routine.
 		 * blocking_notifier_chain_unregister takes the wr lock to do
ipc/util.h +1 -0
Index: b/ipc/util.h
===================================================================
@@ -122,6 +122,7 @@ extern struct msg_msg *load_msg(const vo
 extern int store_msg(void __user *dest, struct msg_msg *msg, int len);
 
 extern void recompute_msgmni(struct ipc_namespace *);
+extern void recompute_msgmnb(struct ipc_namespace *);
 
 static inline int ipc_buildid(int id, int seq)
 {
Signature · 2 lines
Solofo Ramangalahy
Bull SA.
#3 lore raw

RFC 3/6 sysv ipc: do not recompute msgmni anymore if explicitely set by user

In reply to [email protected]

From: Solofo Ramangalahy <[email protected]>

To disable recomputation when user explicitely set a value, reuse the callback defined for msgmni for msgmnb.

As msgmni and msgmnb are correlated, user settings of any of the two disable recomputation of both, for now. This will be refined in a later patch.

---
 ipc/ipc_sysctl.c |    4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)
ipc/ipc_sysctl.c +2 -2
Index: b/ipc/ipc_sysctl.c
===================================================================
@@ -210,8 +210,8 @@ static struct ctl_table ipc_kern_table[]
 		.data		= &init_ipc_ns.msg_ctlmnb,
 		.maxlen		= sizeof (init_ipc_ns.msg_ctlmnb),
 		.mode		= 0644,
-		.proc_handler	= proc_ipc_dointvec,
-		.strategy	= sysctl_ipc_data,
+		.proc_handler	= proc_ipc_callback_dointvec,
+		.strategy	= sysctl_ipc_registered_data,
 	},
 	{
 		.ctl_name	= KERN_SEM,
Signature · 2 lines
Solofo Ramangalahy
Bull SA.
#4 lore raw

RFC 4/6 sysv ipc: re-enable msgmnb automatic recomputing if set to negative

In reply to [email protected]

From: Solofo Ramangalahy <[email protected]>

When a negative value is put in /proc/sys/kernel/msgmnb automatic recomputing is re-enabled.

---
 ipc/ipc_sysctl.c |    1 +
 1 file changed, 1 insertion(+)
ipc/ipc_sysctl.c +1 -0
Index: b/ipc/ipc_sysctl.c
===================================================================
@@ -42,6 +42,7 @@ static void tunable_set_callback(int val
 		 * Re-enable automatic recomputing only if not already
 		 * enabled.
 		 */
+		recompute_msgmnb(current->nsproxy->ipc_ns);
 		recompute_msgmni(current->nsproxy->ipc_ns);
 		cond_register_ipcns_notifier(current->nsproxy->ipc_ns);
 	}
Signature · 2 lines
Solofo Ramangalahy
Bull SA.
#5 lore raw

RFC 5/6 sysv ipc: deconnect msgmnb and msgmni deactivation and reactivation

In reply to [email protected]

From: Solofo Ramangalahy <[email protected]>

The msgmnb and msgmni values are coupled for deactivation and reactivation of value computation.

The uncoupling of msgmn{b,i} for deactivation/reactivation of recomputation adds flexibility and testability.

. Flexibility was discussed during the msgmni series development and
ended up with reactivation by negative value on /proc.

. Testability allows to experiment with the automatic computation of
msgmn{b,i} values. For example, if current algorithm does not fit application needs.

---
 include/linux/ipc_namespace.h |    3 +-
 ipc/ipc_sysctl.c              |   45 ++++++++++++++++++++++++++++++++----------
 ipc/ipcns_notifier.c          |    9 --------
 ipc/msg.c                     |    6 +++++
 4 files changed, 43 insertions(+), 20 deletions(-)
include/linux/ipc_namespace.h +2 -1
Index: b/include/linux/ipc_namespace.h
===================================================================
@@ -34,7 +34,9 @@ struct ipc_namespace {
 
 	int		msg_ctlmax;
 	int		msg_ctlmnb;
+	bool		msg_ctlmnb_activated; /* recompute_msgmnb activation */
 	int		msg_ctlmni;
+	bool		msg_ctlmni_activated; /* recompute_msgmni activation */
 	atomic_t	msg_bytes;
 	atomic_t	msg_hdrs;
 
@@ -53,7 +55,6 @@ extern atomic_t nr_ipc_ns;
 #define INIT_IPC_NS(ns)		.ns		= &init_ipc_ns,
 
 extern int register_ipcns_notifier(struct ipc_namespace *);
-extern int cond_register_ipcns_notifier(struct ipc_namespace *);
 extern int unregister_ipcns_notifier(struct ipc_namespace *);
 extern int ipcns_notify(unsigned long);
 
ipc/msg.c +6 -0
Index: b/ipc/msg.c
===================================================================
@@ -91,6 +91,8 @@ void recompute_msgmni(struct ipc_namespa
 	unsigned long allowed;
 	int nb_ns;
 
+	if (!ns->msg_ctlmni_activated)
+		return;
 	si_meminfo(&i);
 	allowed = (((i.totalram - i.totalhigh) / MSG_MEM_SCALE) * i.mem_unit)
 		/ ns->msg_ctlmnb;
@@ -116,6 +118,8 @@ void recompute_msgmni(struct ipc_namespa
  */
 void recompute_msgmnb(struct ipc_namespace *ns)
 {
+	if (!ns->msg_ctlmnb_activated)
+		return;
 	ns->msg_ctlmnb =
 		min(MSGMNB * num_online_cpus(), MSGMNB * MSG_CPU_SCALE);
 }
@@ -123,6 +127,8 @@ void recompute_msgmnb(struct ipc_namespa
 void msg_init_ns(struct ipc_namespace *ns)
 {
 	ns->msg_ctlmax = MSGMAX;
+	ns->msg_ctlmnb_activated = true;
+	ns->msg_ctlmni_activated = true;
 	recompute_msgmnb(ns);
 
 	recompute_msgmni(ns);
ipc/ipc_sysctl.c +35 -10

68 lines not on this page · raw message

ipc/ipcns_notifier.c +0 -9
Index: b/ipc/ipcns_notifier.c
===================================================================
@@ -65,15 +65,6 @@ int register_ipcns_notifier(struct ipc_n
 	return blocking_notifier_chain_register(&ipcns_chain, &ns->ipcns_nb);
 }
 
-int cond_register_ipcns_notifier(struct ipc_namespace *ns)
-{
-	memset(&ns->ipcns_nb, 0, sizeof(ns->ipcns_nb));
-	ns->ipcns_nb.notifier_call = ipcns_callback;
-	ns->ipcns_nb.priority = IPCNS_CALLBACK_PRI;
-	return blocking_notifier_chain_cond_register(&ipcns_chain,
-							&ns->ipcns_nb);
-}
-
 int unregister_ipcns_notifier(struct ipc_namespace *ns)
 {
 	return blocking_notifier_chain_unregister(&ipcns_chain,
Signature · 2 lines
Solofo Ramangalahy
Bull SA.
#6 lore raw

Re: [RFC -mm 5/6] sysv ipc: deconnect msgmnb and msgmni deactivation and reactivation

Quoted 110 lines · [email protected] wrote:

From: Solofo Ramangalahy <[email protected]>

The msgmnb and msgmni values are coupled for deactivation and reactivation of value computation.

The uncoupling of msgmn{b,i} for deactivation/reactivation of recomputation adds flexibility and testability.

. Flexibility was discussed during the msgmni series development and
ended up with reactivation by negative value on /proc.

. Testability allows to experiment with the automatic computation of
msgmn{b,i} values. For example, if current algorithm does not fit application needs.

---
 include/linux/ipc_namespace.h |    3 +-
 ipc/ipc_sysctl.c              |   45 ++++++++++++++++++++++++++++++++----------
 ipc/ipcns_notifier.c          |    9 --------
 ipc/msg.c                     |    6 +++++
 4 files changed, 43 insertions(+), 20 deletions(-)

Index: b/include/linux/ipc_namespace.h
===================================================================

include/linux/ipc_namespace.h +2 -1
@@ -34,7 +34,9 @@ struct ipc_namespace {
 
 	int		msg_ctlmax;
 	int		msg_ctlmnb;
+	bool		msg_ctlmnb_activated; /* recompute_msgmnb activation */
 	int		msg_ctlmni;
+	bool		msg_ctlmni_activated; /* recompute_msgmni activation */
 	atomic_t	msg_bytes;
 	atomic_t	msg_hdrs;
 
@@ -53,7 +55,6 @@ extern atomic_t nr_ipc_ns;
 #define INIT_IPC_NS(ns)		.ns		= &init_ipc_ns,
 
 extern int register_ipcns_notifier(struct ipc_namespace *);
-extern int cond_register_ipcns_notifier(struct ipc_namespace *);
 extern int unregister_ipcns_notifier(struct ipc_namespace *);
 extern int ipcns_notify(unsigned long);
 

Index: b/ipc/msg.c
===================================================================

ipc/msg.c +6 -0
@@ -91,6 +91,8 @@ void recompute_msgmni(struct ipc_namespa
 	unsigned long allowed;
 	int nb_ns;
 
+	if (!ns->msg_ctlmni_activated)
+		return;
 	si_meminfo(&i);
 	allowed = (((i.totalram - i.totalhigh) / MSG_MEM_SCALE) * i.mem_unit)
 		/ ns->msg_ctlmnb;
@@ -116,6 +118,8 @@ void recompute_msgmni(struct ipc_namespa
  */
 void recompute_msgmnb(struct ipc_namespace *ns)
 {
+	if (!ns->msg_ctlmnb_activated)
+		return;
 	ns->msg_ctlmnb =
 		min(MSGMNB * num_online_cpus(), MSGMNB * MSG_CPU_SCALE);
 }
@@ -123,6 +127,8 @@ void recompute_msgmnb(struct ipc_namespa
 void msg_init_ns(struct ipc_namespace *ns)
 {
 	ns->msg_ctlmax = MSGMAX;
+	ns->msg_ctlmnb_activated = true;
+	ns->msg_ctlmni_activated = true;
 	recompute_msgmnb(ns);
 
 	recompute_msgmni(ns);

Index: b/ipc/ipc_sysctl.c
===================================================================

ipc/ipc_sysctl.c +17 -4
@@ -33,18 +33,42 @@ static void *get_ipc(ctl_table *table)
  * add/remove or ipc namespace creation/removal.
  * They can come back to a recomputable state by being set to a <0 value.
  */
-static void tunable_set_callback(int val)
+static void tunable_set_callback(int val, ctl_table *table)
 {
-	if (val >= 0)
-		unregister_ipcns_notifier(current->nsproxy->ipc_ns);
-	else {
+	int tunable = table->ctl_name;
+
+	if (val >= 0) {
+		switch (tunable) {
+		case KERN_MSGMNB:
+			current->nsproxy->ipc_ns->msg_ctlmnb_activated = false;
+			break;
+		case KERN_MSGMNI:
+			current->nsproxy->ipc_ns->msg_ctlmni_activated = false;
+			break;
+		default:
+			printk(KERN_ERR "ipc: unexpected value %s\n",
+			       table->procname);
+			break;
+		}
+	} else {
 		/*
 		 * Re-enable automatic recomputing only if not already

ipc/ipc_sysctl.c static void *get_ipc(ctl_table *table) · line 62

Quoted 10 lines
ipc/ipc_sysctl.c +5 -3
 		 * enabled.
 		 */
-		recompute_msgmnb(current->nsproxy->ipc_ns);
-		recompute_msgmni(current->nsproxy->ipc_ns);
-		cond_register_ipcns_notifier(current->nsproxy->ipc_ns);
+		switch (tunable) {
+		case KERN_MSGMNB:
+			current->nsproxy->ipc_ns->msg_ctlmnb_activated = true;
+			recompute_msgmnb(current->nsproxy->ipc_ns);
+							/* fall through */

You shouldn't be falling through here: if you do that, re-enablng msgmnb will re-enable msgmni too.

Quoted 44 lines
ipc/ipc_sysctl.c +13 -3
+		case KERN_MSGMNI:
+			current->nsproxy->ipc_ns->msg_ctlmni_activated = true;
+			recompute_msgmni(current->nsproxy->ipc_ns);
+			break;
+		default:
+			printk(KERN_ERR "ipc: unexpected value %s\n",
+			       table->procname);
+			break;
+		}
 	}
 }
 
@@ -72,7 +96,8 @@ static int proc_ipc_callback_dointvec(ct
 	rc = proc_dointvec(&ipc_table, write, filp, buffer, lenp, ppos);
 
 	if (write && !rc && lenp_bef == *lenp)
-		tunable_set_callback(*((int *)(ipc_table.data)));
+		BUG_ON(table == NULL);
+		tunable_set_callback(*((int *)(ipc_table.data)), table);
 
 	return rc;
 }
@@ -148,8 +173,8 @@ static int sysctl_ipc_registered_data(ct
 		 * Tunable has successfully been changed from userland
 		 */
 		int *data = get_ipc(table);
-
-		tunable_set_callback(*data);
+		BUG_ON(table == NULL);
+		tunable_set_callback(*data, table);
 	}
 
 	return rc;

Index: b/ipc/ipcns_notifier.c
===================================================================

ipc/ipcns_notifier.c +0 -2
@@ -65,15 +65,6 @@ int register_ipcns_notifier(struct ipc_n
 	return blocking_notifier_chain_register(&ipcns_chain, &ns->ipcns_nb);
 }
 
-int cond_register_ipcns_notifier(struct ipc_namespace *ns)
-{

ipc/ipcns_notifier.c int register_ipcns_notifier(struct ipc_n · line 70

Quoted 10 lines
ipc/ipcns_notifier.c +0 -7
-	memset(&ns->ipcns_nb, 0, sizeof(ns->ipcns_nb));
-	ns->ipcns_nb.notifier_call = ipcns_callback;
-	ns->ipcns_nb.priority = IPCNS_CALLBACK_PRI;
-	return blocking_notifier_chain_cond_register(&ipcns_chain,
-							&ns->ipcns_nb);
-}
-
 int unregister_ipcns_notifier(struct ipc_namespace *ns)
 {
 	return blocking_notifier_chain_unregister(&ipcns_chain,

Doing this, we are completly loosing the benefits of the notification chains: since the the notifier blocks remain registered + we are unconditionally adding a test at the top of each recompute routine. But the other choice would hve been to define another notifier chain dedicated to msgmnb. I'm not convinced about what is the best solution?

Regards,
Nadia

#7 lore raw

RFC 6/6 sysv ipc: documentation for msgmnb scaling wrt. cpus

In reply to [email protected]

From: Solofo Ramangalahy <[email protected]>

Add documentation to explain how to disable and reenable the computation mechanism.

---

 Documentation/sysctl/kernel.txt |   27 +++++++++++++++++++++++++++
 1 file changed, 27 insertions(+)
Documentation/sysctl/kernel.txt +27 -0
Index: b/Documentation/sysctl/kernel.txt
===================================================================
@@ -179,6 +179,33 @@ kernel stack.
 
 ==============================================================
 
+msgmnb
+
+Maximum size in bytes, not in message count, of a single SystemV IPC
+message queue (b stands for bytes).
+
+This value is dynamic and depends on the online cpu count of the
+machine (taking cpu hotplug into account).
+
+Computed values are between MSGMNB and MSGMNB*MSG_CPU_SCALE #define
+constants (currently [16384,65536]).
+
+The exact value is automatically (re)computed, but:
+. If the value is positioned from user space (via procfs or sysctl()),
+  to a positive value then the automatic recomputation is
+  disabled. This leaves control to user space. E.g.
+
+  # echo 16384 > /proc/sys/kernel/msgmnb
+
+. If the value is positioned from user space to a negative value, then
+  the computation is reenabled. E.g.
+
+  # echo -1 > /proc/sys/kernel/msgmnb
+
+See recompute_msgmnb() function in ipc/ directory for details.
+
+==============================================================
+
 osrelease, ostype & version:
 
 # cat osrelease
Signature · 2 lines
Solofo Ramangalahy
Bull SA.
#8 lore raw
In reply to [email protected]
Quoted 2 lines · On Friday 06 June 2008 16:09, [email protected] wrote:

The size in bytes of a SysV IPC message queue, msgmnb, is too small for large machines, but we don't want to bloat small machines

What's your evidence for this? Can you provide before / after performance numbers?

Also, when scaling things like this, it is probably more usual to use a log scale rather than linear, so that's a thought.

Quoted 67 lines

Several methods are used already to modify (mainly increase) msgmnb:
. distribution specific patch
. system wide sysctl.conf
. application specific tuning via /proc/sys/kernel/msgmnb

Integrating this series would:
. reflect hardware and software evolutions and diversity, . reduce configuration/tuning for the applications.

Here is the timeline of the evolution of MSG* #defines:
Year 1994 1999 1999 2008
Version 1.0 2.3.27 2.3.30 2.6.24
#define MSGMNI 128 128 16 16
#define MSGMAX 4056 8192 8192 8192
#define MSGMNB 16384 16384 16384 16384

This patch series scales msgmnb, with respect to the number of cpus/cores for larger machines. For uniprocessor machines the value does not increase.

This series is similar to (and depends on) the series which scales msgmni, the number of IPC message queue identifiers, to the amount of low memory.
While Nadia's previous series scaled msgmni along the memory axis, hence the message pool (msgmni x msgmnb), this series uses a second
axis: the number of online CPUs.
As well as covering the (cpu,memory) space of machines size, this reflects the parallelism allowed by lockless send/receive for in-flight messages in queues (msgmnb / msgmax messages).

The initial scaling is done at initialization of the ipc namespace. Furthermore, the value becomes dynamic with respect to cpu hotplug.

The msgmni and msgmnb values become dependent, as the value of msgmni is computed with respect to the value of msgmnb.

The series is as follows:
. patch 1 introduces the scaling function
. patch 2 deals with cpu hotplug
. patch 3 allows user space to disable the scaling mechanism . patch 4 allows user space to reenable the scaling mechanism . patch 5 finer grain disabling/reenabling scaling mechanism
(disconnect msgmnb and msgmni)
. patch 6 adds documentation

---

The series applies to 2.6.26-rc2-mm1 + patch suppressing KERN_INFO messages as discussed at:
http://article.gmane.org/gmane.linux.kernel/686229
"[PATCH 1/1] Only output msgmni value at boot time"
(in mmotm: ipc-only-output-msgmni-value-at-boot-time.patch)

The plan would be to have this ready for the 2.6.27 merge window if there are no objections.

 Documentation/sysctl/kernel.txt |   27 ++++++++++++++++++++++
 include/linux/ipc_namespace.h   |    4 ++-
 include/linux/msg.h             |    5 ++++
 ipc/ipc_sysctl.c                |   48
++++++++++++++++++++++++++++++---------- ipc/ipcns_notifier.c            | 
 23 +++++++------------
 ipc/msg.c                       |   25 +++++++++++++++++---
 ipc/util.c                      |   28 +++++++++++++++++++++++
 ipc/util.h                      |    1
 8 files changed, 131 insertions(+), 30 deletions(-)
#9 lore raw

Hi Nick,

Quoted 6 lines · Nick Piggin
Quoted 2 lines · On Friday 06 June 2008 16:09, [email protected] wrote:

The size in bytes of a SysV IPC message queue, msgmnb, is too small for large machines, but we don't want to bloat small machines

What's your evidence for this? Can you provide before / after performance numbers?

Maybe I have not been clear enough that this is not directly about performance, but more changing default value. So maybe "scale" in the title is misleading.

The evidence would be that these default values are changed either by a patch or "manually":

Quoted 4 lines
Quoted 4 lines

Several methods are used already to modify (mainly increase) msgmnb:
. distribution specific patch
. system wide sysctl.conf
. application specific tuning via /proc/sys/kernel/msgmnb

Further "evidence" could be googling for "linux msgmnb 65536", to see that tuning for benchmarks or recommended application configuration increase the value.

This is just settings default values. Performance test results would not be different from those obtained by setting the values before running the tests.

So here :

Quoted 6 lines
Quoted 6 lines

Here is the timeline of the evolution of MSG* #defines:
Year 1994 1999 1999 2008
Version 1.0 2.3.27 2.3.30 2.6.24
#define MSGMNI 128 128 16 16
#define MSGMAX 4056 8192 8192 8192
#define MSGMNB 16384 16384 16384 16384

I have 65536 instead of 16384 for msgmnb
(1982 instead of 16 for msgmni)
for my 4 cpus/4GB x86_64 machine.

Some result with pmsg used in recent discussions about performance gives:
16384/16:
./pmsg 4 10 |grep Total
Total: 9795993

65536/1982:
./pmsg 4 10 |grep Total
Total: 9829590

Quoted 2 lines

Also, when scaling things like this, it is probably more usual to use a log scale rather than linear, so that's a thought.

Agreed, in general.
Here, there are only 4 values, so I do not think it is worth using a log scale.
If different values are desirable (finer grain, bigger,...), then the formula can be easily refined:
min(MSGMNB * num_online_cpus(), MSGMNB * MSG_CPU_SCALE); What I did for the formula is simply taking the old value, a known modified value and the intermediate values.

I hope this answers your questions,

Signature · 1 line
solofo
#10 lore raw
In reply to [email protected]
Quoted 39 lines · [email protected] wrote:

The size in bytes of a SysV IPC message queue, msgmnb, is too small for large machines, but we don't want to bloat small machines

Several methods are used already to modify (mainly increase) msgmnb:
. distribution specific patch
. system wide sysctl.conf
. application specific tuning via /proc/sys/kernel/msgmnb

Integrating this series would:
. reflect hardware and software evolutions and diversity, . reduce configuration/tuning for the applications.

Here is the timeline of the evolution of MSG* #defines:
Year 1994 1999 1999 2008
Version 1.0 2.3.27 2.3.30 2.6.24
#define MSGMNI 128 128 16 16
#define MSGMAX 4056 8192 8192 8192
#define MSGMNB 16384 16384 16384 16384

This patch series scales msgmnb, with respect to the number of cpus/cores for larger machines. For uniprocessor machines the value does not increase.

This series is similar to (and depends on) the series which scales msgmni, the number of IPC message queue identifiers, to the amount of low memory.
While Nadia's previous series scaled msgmni along the memory axis, hence the message pool (msgmni x msgmnb), this series uses a second
axis: the number of online CPUs.
As well as covering the (cpu,memory) space of machines size, this reflects the parallelism allowed by lockless send/receive for in-flight messages in queues (msgmnb / msgmax messages).

The initial scaling is done at initialization of the ipc namespace. Furthermore, the value becomes dynamic with respect to cpu hotplug.

The msgmni and msgmnb values become dependent, as the value of msgmni is computed with respect to the value of msgmnb.

Quoted 8 lines

The series is as follows:
. patch 1 introduces the scaling function . patch 2 deals with cpu hotplug . patch 3 allows user space to disable the scaling mechanism
. patch 4 allows user space to reenable the scaling mechanism . patch 5 finer grain disabling/reenabling scaling mechanism
(disconnect msgmnb and msgmni)
. patch 6 adds documentation

Solofo,

Patches 3 and 4 are useless imho. If you really really want to keep them, you should at least merge them.

Regards,
Nadia