public inbox for linux-kernel@vger.kernel.org
 help / color / mirror / Atom feed
From: Peter Zijlstra <peterz@infradead.org>
To: Mike Galbraith <efault@gmx.de>
Cc: Suresh Siddha <suresh.b.siddha@intel.com>,
	Ingo Molnar <mingo@elte.hu>, Venki Pallipadi <venki@google.com>,
	Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com>,
	linux-kernel <linux-kernel@vger.kernel.org>,
	Tim Chen <tim.c.chen@linux.jf.intel.com>,
	alex.shi@intel.com
Subject: Re: [patch v3 5/6] sched, ttwu_queue: queue remote wakeups only when crossing cache domains
Date: Wed, 07 Dec 2011 17:23:47 +0100	[thread overview]
Message-ID: <1323275027.32012.114.camel@twins> (raw)
In-Reply-To: <1322796864.4755.5.camel@marge.simson.net>

On Fri, 2011-12-02 at 04:34 +0100, Mike Galbraith wrote:
> On Thu, 2011-12-01 at 17:07 -0800, Suresh Siddha wrote:
> > plain text document attachment
> > (use_ttwu_queue_when_crossing_cache_domains.patch)
> > From: Mike Galbraith <efault@gmx.de>
> > 
> > Context-switch intensive microbenchmark on a 8-socket system had
> > ~600K times more resched IPI's on each logical CPU because of the
> > TTWU_QUEUE sched feature, which queues the task on the remote cpu's
> > queue and completes the wakeup locally using an IPI.
> > 
> > As the TTWU_QUEUE sched feature is for minimizing the cache-misses
> > associated with the remote wakeups, use the IPI only when the local and
> > the remote cpu's are from different cache domains. Otherwise use the
> > traditional remote wakeup.
> 
> FYI, Peter has already (improved and) queued this patch.

In fact, Ingo (rightfully) refused to take this due to the x86 specific
code in scheduler guts..

Initially the idea was to provide a new arch interface and a fallback
and do the Kconfig thing etc. After a bit of thought I decided against
that for we already have that information in the sched_domain tree
anyway and it should be a simple matter of representing things
differently.

This led to the below patch, which seems to boot on my box. I still hate
the sd_top_spr* names but whatever.. ;-)

---
 kernel/sched/core.c  |   36 +++++++++++++++++++++++++++++++++++-
 kernel/sched/fair.c  |   24 +-----------------------
 kernel/sched/sched.h |   42 ++++++++++++++++++++++++++++++++++++------
 3 files changed, 72 insertions(+), 30 deletions(-)
Index: linux-2.6/kernel/sched/core.c
===================================================================
--- linux-2.6.orig/kernel/sched/core.c
+++ linux-2.6/kernel/sched/core.c
@@ -1511,6 +1511,12 @@ static int ttwu_activate_remote(struct t
 
 }
 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
+
+static inline int ttwu_share_cache(int this_cpu, int that_cpu)
+{
+	return per_cpu(sd_top_spr_id, this_cpu) ==
+		per_cpu(sd_top_spr_id, that_cpu);
+}
 #endif /* CONFIG_SMP */
 
 static void ttwu_queue(struct task_struct *p, int cpu)
@@ -1518,7 +1524,7 @@ static void ttwu_queue(struct task_struc
 	struct rq *rq = cpu_rq(cpu);
 
 #if defined(CONFIG_SMP)
-	if (sched_feat(TTWU_QUEUE) && cpu != smp_processor_id()) {
+	if (sched_feat(TTWU_QUEUE) && !ttwu_share_cache(smp_processor_id(), cpu)) {
 		sched_clock_cpu(cpu); /* sync clocks x-cpu */
 		ttwu_queue_remote(p, cpu);
 		return;
@@ -5751,6 +5757,32 @@ static void destroy_sched_domains(struct
 }
 
 /*
+ * Keep a special pointer to the highest sched_domain that has
+ * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this
+ * allows us to avoid some pointer chasing select_idle_sibling().
+ *
+ * Also keep a unique ID per domain (we use the first cpu number in
+ * the cpumask of the domain), this allows us to quickly tell if
+ * two cpus are in the same cache domain, see ttwu_share_cache().
+ */
+DEFINE_PER_CPU(struct sched_domain *, sd_top_spr);
+DEFINE_PER_CPU(int, sd_top_spr_id);
+
+static void update_top_cache_domain(int cpu)
+{
+	struct sched_domain *sd;
+	int id = -1;
+
+	sd = highest_flag_domain(cpu, SD_SHARE_PKG_RESOURCES);
+	if (sd)
+		id = cpumask_first(sched_domain_span(sd));
+
+
+	rcu_assign_pointer(per_cpu(sd_top_spr, cpu), sd);
+	per_cpu(sd_top_spr_id, cpu) = id;
+}
+
+/*
  * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
  * hold the hotplug lock.
  */
@@ -5789,6 +5821,8 @@ cpu_attach_domain(struct sched_domain *s
 	tmp = rq->sd;
 	rcu_assign_pointer(rq->sd, sd);
 	destroy_sched_domains(tmp, cpu);
+
+	update_top_cache_domain(cpu);
 }
 
 /* cpus with isolated domains */
Index: linux-2.6/kernel/sched/fair.c
===================================================================
--- linux-2.6.orig/kernel/sched/fair.c
+++ linux-2.6/kernel/sched/fair.c
@@ -2644,28 +2644,6 @@ find_idlest_cpu(struct sched_group *grou
 	return idlest;
 }
 
-/**
- * highest_flag_domain - Return highest sched_domain containing flag.
- * @cpu:	The cpu whose highest level of sched domain is to
- *		be returned.
- * @flag:	The flag to check for the highest sched_domain
- *		for the given cpu.
- *
- * Returns the highest sched_domain of a cpu which contains the given flag.
- */
-static inline struct sched_domain *highest_flag_domain(int cpu, int flag)
-{
-	struct sched_domain *sd, *hsd = NULL;
-
-	for_each_domain(cpu, sd) {
-		if (!(sd->flags & flag))
-			break;
-		hsd = sd;
-	}
-
-	return hsd;
-}
-
 /*
  * Try and locate an idle CPU in the sched_domain.
  */
@@ -2696,7 +2674,7 @@ static int select_idle_sibling(struct ta
 	 */
 	rcu_read_lock();
 
-	sd = highest_flag_domain(target, SD_SHARE_PKG_RESOURCES);
+	sd = rcu_dereference(per_cpu(sd_top_spr, target));
 	for_each_lower_domain(sd) {
 		sg = sd->groups;
 		do {
Index: linux-2.6/kernel/sched/sched.h
===================================================================
--- linux-2.6.orig/kernel/sched/sched.h
+++ linux-2.6/kernel/sched/sched.h
@@ -487,6 +487,14 @@ static inline int cpu_of(struct rq *rq)
 
 DECLARE_PER_CPU(struct rq, runqueues);
 
+#define cpu_rq(cpu)		(&per_cpu(runqueues, (cpu)))
+#define this_rq()		(&__get_cpu_var(runqueues))
+#define task_rq(p)		cpu_rq(task_cpu(p))
+#define cpu_curr(cpu)		(cpu_rq(cpu)->curr)
+#define raw_rq()		(&__raw_get_cpu_var(runqueues))
+
+#ifdef CONFIG_SMP
+
 #define rcu_dereference_check_sched_domain(p) \
 	rcu_dereference_check((p), \
 			      lockdep_is_held(&sched_domains_mutex))
@@ -499,15 +507,37 @@ DECLARE_PER_CPU(struct rq, runqueues);
  * preempt-disabled sections.
  */
 #define for_each_domain(cpu, __sd) \
-	for (__sd = rcu_dereference_check_sched_domain(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
+	for (__sd = rcu_dereference_check_sched_domain(cpu_rq(cpu)->sd); \
+			__sd; __sd = __sd->parent)
 
 #define for_each_lower_domain(sd) for (; sd; sd = sd->child)
 
-#define cpu_rq(cpu)		(&per_cpu(runqueues, (cpu)))
-#define this_rq()		(&__get_cpu_var(runqueues))
-#define task_rq(p)		cpu_rq(task_cpu(p))
-#define cpu_curr(cpu)		(cpu_rq(cpu)->curr)
-#define raw_rq()		(&__raw_get_cpu_var(runqueues))
+/**
+ * highest_flag_domain - Return highest sched_domain containing flag.
+ * @cpu:	The cpu whose highest level of sched domain is to
+ *		be returned.
+ * @flag:	The flag to check for the highest sched_domain
+ *		for the given cpu.
+ *
+ * Returns the highest sched_domain of a cpu which contains the given flag.
+ */
+static inline struct sched_domain *highest_flag_domain(int cpu, int flag)
+{
+	struct sched_domain *sd, *hsd = NULL;
+
+	for_each_domain(cpu, sd) {
+		if (!(sd->flags & flag))
+			break;
+		hsd = sd;
+	}
+
+	return hsd;
+}
+
+DECLARE_PER_CPU(struct sched_domain *, sd_top_spr);
+DECLARE_PER_CPU(int, sd_top_spr_id);
+
+#endif /* CONFIG_SMP */
 
 #include "stats.h"
 #include "auto_group.h"


  reply	other threads:[~2011-12-07 16:24 UTC|newest]

Thread overview: 30+ messages / expand[flat|nested]  mbox.gz  Atom feed  top
2011-12-02  1:07 [patch v3 0/6] nohz idle load balancing patches Suresh Siddha
2011-12-02  1:07 ` [patch v3 1/6] sched, nohz: introduce nohz_flags in the struct rq Suresh Siddha
2011-12-06  9:53   ` [tip:sched/core] sched, nohz: Introduce nohz_flags in 'struct rq' tip-bot for Suresh Siddha
2011-12-06 12:14   ` [patch v3 1/6] sched, nohz: introduce nohz_flags in the struct rq Srivatsa Vaddagiri
2011-12-06 19:26     ` Suresh Siddha
2011-12-06 19:39       ` Peter Zijlstra
2011-12-06 20:24       ` [tip:sched/core] sched, nohz: Set the NOHZ_BALANCE_KICK flag for idle load balancer tip-bot for Suresh Siddha
2011-12-02  1:07 ` [patch v3 2/6] sched, nohz: track nr_busy_cpus in the sched_group_power Suresh Siddha
2011-12-06  9:54   ` [tip:sched/core] sched, nohz: Track " tip-bot for Suresh Siddha
2011-12-02  1:07 ` [patch v3 3/6] sched, nohz: sched group, domain aware nohz idle load balancing Suresh Siddha
2011-12-06  6:37   ` Srivatsa Vaddagiri
2011-12-06 19:19     ` Suresh Siddha
2011-12-06 20:24       ` [tip:sched/core] sched, nohz: Fix the idle cpu check in nohz_idle_balance tip-bot for Suresh Siddha
     [not found]     ` <A75BCAD09CE00A4280CDD4429D85F1F9261B42A1F9@orsmsx501.amr.corp.intel.com>
2011-12-06 19:27       ` [patch v3 3/6] sched, nohz: sched group, domain aware nohz idle load balancing Suresh Siddha
2011-12-06  9:54   ` [tip:sched/core] sched, nohz: Implement " tip-bot for Suresh Siddha
2011-12-02  1:07 ` [patch v3 4/6] sched, nohz: cleanup the find_new_ilb() using sched groups nr_busy_cpus Suresh Siddha
2011-12-06  9:55   ` [tip:sched/core] sched, nohz: Clean up " tip-bot for Suresh Siddha
2011-12-02  1:07 ` [patch v3 5/6] sched, ttwu_queue: queue remote wakeups only when crossing cache domains Suresh Siddha
2011-12-02  3:34   ` Mike Galbraith
2011-12-07 16:23     ` Peter Zijlstra [this message]
2011-12-07 19:20       ` Suresh Siddha
2011-12-08  6:06         ` Mike Galbraith
2011-12-08  9:41           ` Peter Zijlstra
2011-12-08  9:29         ` Peter Zijlstra
2011-12-08 19:34           ` Suresh Siddha
2011-12-08 21:50             ` Peter Zijlstra
2011-12-08 21:51               ` Peter Zijlstra
2011-12-08 10:02         ` Peter Zijlstra
2011-12-21 11:41           ` [tip:sched/core] sched: Only queue remote wakeups when crossing cache boundaries tip-bot for Peter Zijlstra
2011-12-02  1:07 ` [patch v3 6/6] sched: fix the sched group node allocation for SD_OVERLAP domain Suresh Siddha

Reply instructions:

You may reply publicly to this message via plain-text email
using any one of the following methods:

* Save the following mbox file, import it into your mail client,
  and reply-to-all from there: mbox

  Avoid top-posting and favor interleaved quoting:
  https://en.wikipedia.org/wiki/Posting_style#Interleaved_style

* Reply using the --to, --cc, and --in-reply-to
  switches of git-send-email(1):

  git send-email \
    --in-reply-to=1323275027.32012.114.camel@twins \
    --to=peterz@infradead.org \
    --cc=alex.shi@intel.com \
    --cc=efault@gmx.de \
    --cc=linux-kernel@vger.kernel.org \
    --cc=mingo@elte.hu \
    --cc=suresh.b.siddha@intel.com \
    --cc=tim.c.chen@linux.jf.intel.com \
    --cc=vatsa@linux.vnet.ibm.com \
    --cc=venki@google.com \
    /path/to/YOUR_REPLY

  https://kernel.org/pub/software/scm/git/docs/git-send-email.html

* If your mail client supports setting the In-Reply-To header
  via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line before the message body.
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox