[Xen-devel] [PATCH v2 6/7] xen: credit1: treat pCPUs more evenly during balancing.

[Dario Faggioli <dario.faggioli@xxxxxxxxxx>]

Right now, we use cpumask_first() for going through
the bus pCPUs in csched_load_balance(). This means
not all pCPUs have equal chances of seeing their
pending work stolen. It also means there is more
runqueue lock pressure on lower ID pCPUs.

To avoid all this, let's record and remember, for
each NUMA node, from what pCPU we have stolen for
last, and start from that the following time.

Signed-off-by: Dario Faggioli <dario.faggioli@xxxxxxxxxx>
---
Cc: George Dunlap <george.dunlap@xxxxxxxxxxxxx>
Cc: Anshul Makkar <anshul.makkar@xxxxxxxxxx>
---
 xen/common/sched_credit.c |   37 +++++++++++++++++++++++++++++++++----
 1 file changed, 33 insertions(+), 4 deletions(-)

diff --git a/xen/common/sched_credit.c b/xen/common/sched_credit.c
index 09e3192..c37a65c 100644
--- a/xen/common/sched_credit.c
+++ b/xen/common/sched_credit.c
@@ -229,6 +229,7 @@ struct csched_private {
     uint32_t credit;
     int credit_balance;
     uint32_t runq_sort;
+    uint32_t *balance_bias;
     unsigned ratelimit_us;
     /* Period of master and tick in milliseconds */
     unsigned tslice_ms, tick_period_us, ticks_per_tslice;
@@ -547,6 +548,7 @@ csched_deinit_pdata(const struct scheduler *ops, void 
*pcpu, int cpu)
 {
     struct csched_private *prv = CSCHED_PRIV(ops);
     struct csched_pcpu *spc = pcpu;
+    unsigned int node = cpu_to_node(cpu);
     unsigned long flags;
 
     /*
@@ -570,6 +572,12 @@ csched_deinit_pdata(const struct scheduler *ops, void 
*pcpu, int cpu)
         prv->master = cpumask_first(prv->cpus);
         migrate_timer(&prv->master_ticker, prv->master);
     }
+    if ( prv->balance_bias[node] == cpu )
+    {
+        cpumask_and(cpumask_scratch, prv->cpus, &node_to_cpumask(node));
+        if ( !cpumask_empty(cpumask_scratch) )
+            prv->balance_bias[node] =  cpumask_first(cpumask_scratch);
+    }
     kill_timer(&spc->ticker);
     if ( prv->ncpus == 0 )
         kill_timer(&prv->master_ticker);
@@ -609,6 +617,10 @@ init_pdata(struct csched_private *prv, struct csched_pcpu 
*spc, int cpu)
                   NOW() + MILLISECS(prv->tslice_ms));
     }
 
+    cpumask_and(cpumask_scratch, prv->cpus, 
&node_to_cpumask(cpu_to_node(cpu)));
+    if ( cpumask_weight(cpumask_scratch) == 1 )
+        prv->balance_bias[cpu_to_node(cpu)] = cpu;
+
     init_timer(&spc->ticker, csched_tick, (void *)(unsigned long)cpu, cpu);
     set_timer(&spc->ticker, NOW() + MICROSECS(prv->tick_period_us) );
 
@@ -1710,7 +1722,7 @@ csched_load_balance(struct csched_private *prv, int cpu,
     struct csched_vcpu *speer;
     cpumask_t workers;
     cpumask_t *online;
-    int peer_cpu, peer_node, bstep;
+    int peer_cpu, first_cpu, peer_node, bstep;
     int node = cpu_to_node(cpu);
 
     BUG_ON( cpu != snext->vcpu->processor );
@@ -1754,9 +1766,10 @@ csched_load_balance(struct csched_private *prv, int cpu,
             cpumask_and(&workers, &workers, &node_to_cpumask(peer_node));
             __cpumask_clear_cpu(cpu, &workers);
 
-            peer_cpu = cpumask_first(&workers);
-            if ( peer_cpu >= nr_cpu_ids )
+            first_cpu = cpumask_cycle(prv->balance_bias[peer_node], &workers);
+            if ( first_cpu >= nr_cpu_ids )
                 goto next_node;
+            peer_cpu = first_cpu;
             do
             {
                 spinlock_t *lock;
@@ -1820,13 +1833,19 @@ csched_load_balance(struct csched_private *prv, int cpu,
                 if ( speer != NULL )
                 {
                     *stolen = 1;
+                    /*
+                     * Next time we'll look for work to steal on this node, we
+                     * will start from the next pCPU, with respect to this one,
+                     * so we don't risk stealing always from the same ones.
+                     */
+                    prv->balance_bias[peer_node] = peer_cpu;
                     return speer;
                 }
 
  next_cpu:
                 peer_cpu = cpumask_cycle(peer_cpu, &workers);
 
-            } while( peer_cpu != cpumask_first(&workers) );
+            } while( peer_cpu != first_cpu );
 
  next_node:
             peer_node = cycle_node(peer_node, node_online_map);
@@ -2178,10 +2197,19 @@ csched_init(struct scheduler *ops)
     prv = xzalloc(struct csched_private);
     if ( prv == NULL )
         return -ENOMEM;
+
+    prv->balance_bias = xzalloc_array(uint32_t, MAX_NUMNODES);
+    if ( prv->balance_bias == NULL )
+    {
+        xfree(prv);
+        return -ENOMEM;
+    }
+
     if ( !zalloc_cpumask_var(&prv->cpus) ||
          !zalloc_cpumask_var(&prv->idlers) )
     {
         free_cpumask_var(prv->cpus);
+        xfree(prv->balance_bias);
         xfree(prv);
         return -ENOMEM;
     }
@@ -2227,6 +2255,7 @@ csched_deinit(struct scheduler *ops)
         ops->sched_data = NULL;
         free_cpumask_var(prv->cpus);
         free_cpumask_var(prv->idlers);
+        xfree(prv->balance_bias);
         xfree(prv);
     }
 }


_______________________________________________
Xen-devel mailing list
Xen-devel@xxxxxxxxxxxxx
https://lists.xen.org/xen-devel