linux-kernel - [PATCH v2] sched/fair: Fix asym packing to select correct cpu

lists.openwall.net		lists / announce owl-users owl-dev john-users john-dev passwdqc-users yescrypt popa3d-users / oss-security kernel-hardening musl sabotage tlsify passwords / crypt-dev xvendor / Bugtraq Full-Disclosure linux-kernel linux-netdev linux-ext4 linux-hardening linux-cve-announce PHC
Open Source and information security mailing list archives

Hash Suite: Windows password security audit tool. GUI, reports in PDF.

[<prev] [next>] [thread-next>] [day] [month] [year] [list]

Message-Id: <1459948660-16073-1-git-send-email-srikar@linux.vnet.ibm.com>
Date:	Wed,  6 Apr 2016 18:47:40 +0530
From:	Srikar Dronamraju <srikar@...ux.vnet.ibm.com>
To:	Ingo Molnar <mingo@...nel.org>,
	Peter Zijlstra <peterz@...radead.org>
Cc:	linux-kernel@...r.kernel.org, srikar@...ux.vnet.ibm.com,
	Gautham R Shenoy <ego@...ux.vnet.ibm.com>,
	Michael Neuling <mikey@...ling.org>,
	Vaidyanathan Srinivasan <svaidy@...ux.vnet.ibm.com>
Subject: [PATCH v2] sched/fair: Fix asym packing to select correct cpu

When asymmetric packing is set in the sched_domain and target cpu is
busy, update_sd_pick_busiest() may not select the busiest runqueue.
When target cpu is busy, find_busiest_group() will ignore checks for
asym packing and may continue to load balance using the currently
selected not-the-busiest runqueue as source runqueue.
Selecting the busiest runqueue as source when the target cpu is busy,
should result in achieving much better load balance.

Also when target cpu is not busy and asymmetric packing is set in sd,
select higher cpu as source cpu for load balancing.

While doing this change, move the check to see if target cpu is busy
into check_asym_packing().

The extent of performance benefit from this change decreases with the
increasing load. However there is benefit in undercommit as well as
overcommit conditions.

1. Record per second ebizzy (32 threads) on a 64 cpu power 7 box. (5 iterations)
4.6.0-rc2
	Testcase:         Min         Max         Avg      StdDev
	  ebizzy:  5223767.00 10368236.00  7946971.00  1753094.76

4.6.0-rc2+asym-changes
	Testcase:         Min         Max         Avg      StdDev     %Change
	  ebizzy:  8617191.00 13872356.00 11383980.00  1783400.89     +24.78%

2. Record per second ebizzy (64 threads) on a 64 cpu power 7 box. (5 iterations)
4.6.0-rc2
	Testcase:         Min         Max         Avg      StdDev
	  ebizzy:  6497666.00 18399783.00 10818093.20  4051452.08

4.6.0-rc2+asym-changes
	Testcase:         Min         Max         Avg      StdDev     %Change
	  ebizzy:  7567365.00 19456937.00 11674063.60  4295407.48      +4.40%

3. Record per second ebizzy (128 threads) on a 64 cpu power 7 box. (5 iterations)
4.6.0-rc2
	Testcase:         Min         Max         Avg      StdDev
	  ebizzy: 37073983.00 40341911.00 38776241.80  1259766.82

4.6.0-rc2+asym-changes
	Testcase:         Min         Max         Avg      StdDev     %Change
	  ebizzy: 38030399.00 41333378.00 39827404.40  1255001.86      +2.54%

Signed-off-by: Srikar Dronamraju <srikar@...ux.vnet.ibm.com>
---
since v1: Rebased to 4.6-rc2
	: Updated patch description
        : Formatted performance numbers as suggested by Michael Neuling.
	: Added comments as suggested by Peter Zijlstra.
	v1 link:  http://mid.gmane.org/1458732880-28382-1-git-send-email-srikar@linux.vnet.ibm.com

 kernel/sched/fair.c | 12 +++++++++---
 1 file changed, 9 insertions(+), 3 deletions(-)

diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c
index 0fe30e6..b09c67a 100644
--- a/kernel/sched/fair.c
+++ b/kernel/sched/fair.c
@@ -6626,6 +6626,9 @@ static bool update_sd_pick_busiest(struct lb_env *env,
 	if (!(env->sd->flags & SD_ASYM_PACKING))
 		return true;
 
+	/* No ASYM_PACKING if target cpu is already busy */
+	if (env->idle == CPU_NOT_IDLE)
+		return true;
 	/*
 	 * ASYM_PACKING needs to move all the work to the lowest
 	 * numbered CPUs in the group, therefore mark all groups
@@ -6635,7 +6638,8 @@ static bool update_sd_pick_busiest(struct lb_env *env,
 		if (!sds->busiest)
 			return true;
 
-		if (group_first_cpu(sds->busiest) > group_first_cpu(sg))
+		/* Prefer to move from highest possible cpu's work */
+		if (group_first_cpu(sds->busiest) < group_first_cpu(sg))
 			return true;
 	}
 
@@ -6781,6 +6785,9 @@ static int check_asym_packing(struct lb_env *env, struct sd_lb_stats *sds)
 	if (!(env->sd->flags & SD_ASYM_PACKING))
 		return 0;
 
+	if (env->idle == CPU_NOT_IDLE)
+		return 0;
+
 	if (!sds->busiest)
 		return 0;
 
@@ -6973,8 +6980,7 @@ static struct sched_group *find_busiest_group(struct lb_env *env)
 	busiest = &sds.busiest_stat;
 
 	/* ASYM feature bypasses nice load balance check */
-	if ((env->idle == CPU_IDLE || env->idle == CPU_NEWLY_IDLE) &&
-	    check_asym_packing(env, &sds))
+	if (check_asym_packing(env, &sds))
 		return sds.busiest;
 
 	/* There is no busy sibling group to pull tasks from */
-- 
1.8.3.1