src/patches/suse-2.6.27.25/patches.suse/sysctl-add-affinity_load_balancing

   1 Subject: add affinity_load_balancing sysctl
   2 From: kraxel@suse.de
   3 References: 176738
   4
   5 Add a sysctl to tweak how the kernel initially schedules threads to cpus.
   6
   7 By default the kernel tries to keep threads on the local cpu (and local
   8 node on NUMA machines).  Depending on the application this may not deliver
   9 the best performance, especially applications with a large working set for
  10 each thread tend to perform better when being scheduled to different nodes
  11 because they can use caches of multiple nodes then.
  12
  13 With this sysctl enabled the kernel will spread threads over the cpus given
  14 and doesn't try to keep them local.
  15
  16 usage:
  17   - set sysctl kernel.affinity_load_balancing = 1
  18   - use taskset or numactl to specify which cpus your task should be
  19     scheduled on.
  20
  21 ---
  22  kernel/sched.c  |   28 ++++++++++++++++++++++++++++
  23  kernel/sysctl.c |   12 ++++++++++++
  24  2 files changed, 40 insertions(+)
  25
  26 --- a/kernel/sched.c
  27 +++ b/kernel/sched.c
  28 @@ -2107,6 +2107,28 @@ find_idlest_cpu(struct sched_group *grou
  29         return idlest;
  30  }
  31
  32 +static int
  33 +find_idlest_cpu_nodomain(struct task_struct *p, int this_cpu)
  34 +{
  35 +       cpumask_t tmp;
  36 +       unsigned long load, min_load = ULONG_MAX;
  37 +       int idlest = -1;
  38 +       int i;
  39 +
  40 +       /* Traverse only the allowed CPUs */
  41 +       cpus_and(tmp, cpu_online_map, p->cpus_allowed);
  42 +
  43 +       for_each_cpu_mask(i, tmp) {
  44 +               load = target_load(i, 1);
  45 +
  46 +               if (load < min_load) {
  47 +                       min_load = load;
  48 +                       idlest = i;
  49 +               }
  50 +       }
  51 +       return idlest;
  52 +}
  53 +
  54  /*
  55   * sched_balance_self: balance the current task (running on cpu) in domains
  56   * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
  57 @@ -2118,11 +2140,17 @@ find_idlest_cpu(struct sched_group *grou
  58   *
  59   * preempt must be disabled.
  60   */
  61 +
  62 +int affinity_load_balancing = 0;
  63 +
  64  static int sched_balance_self(int cpu, int flag)
  65  {
  66         struct task_struct *t = current;
  67         struct sched_domain *tmp, *sd = NULL;
  68
  69 +       if (affinity_load_balancing && !cpus_full(t->cpus_allowed))
  70 +               return find_idlest_cpu_nodomain(t, cpu);
  71 +
  72         for_each_domain(cpu, tmp) {
  73                 /*
  74                  * If power savings logic is enabled for a domain, stop there.
  75 --- a/kernel/sysctl.c
  76 +++ b/kernel/sysctl.c
  77 @@ -237,6 +237,8 @@ static int min_wakeup_granularity_ns;
  78  static int max_wakeup_granularity_ns = NSEC_PER_SEC;   /* 1 second */
  79  #endif
  80
  81 +extern int affinity_load_balancing;
  82 +
  83  static struct ctl_table kern_table[] = {
  84  #ifdef CONFIG_SCHED_DEBUG
  85         {
  86 @@ -872,6 +874,16 @@ static struct ctl_table kern_table[] = {
  87   * NOTE: do not add new entries to this table unless you have read
  88   * Documentation/sysctl/ctl_unnumbered.txt
  89   */
  90 +#ifdef CONFIG_SMP
  91 +       {
  92 +               .ctl_name       = CTL_UNNUMBERED,
  93 +               .procname       = "affinity_load_balancing",
  94 +               .data           = &affinity_load_balancing,
  95 +               .maxlen         = sizeof(affinity_load_balancing),
  96 +               .mode           = 0644,
  97 +               .proc_handler   = &proc_dointvec,
  98 +       },
  99 +#endif
 100         { .ctl_name = 0 }
 101  };
 102