Currently, pmd_rebalance_dry_run() calculate overall variance of
all pmds regardless of their numa location. The overall result may
hide un-balance in an individual numa.
Considering the following case. Numa0 is free because VMs on numa0
are not sending pkts, while numa1 is busy. Within numa1, pmds
workloads are not balanced. Obviously, moving 500 kpps workloads from
pmd 126 to pmd 62 will make numa1 much more balance. For numa1
the variance improment will be almost 100%, because after rebalance
each pmd in numa1 holds same workload(variance ~= 0). But the overall
variance improvement is only about 20%, which may not tigger auto_lb.
```
numa_id core_id kpps
030 0
031 0
094 0
095 0
1 126 1500
1 127 1000
163 1000
162 500
```
As auto_lb doesn't balance workload across numa nodes. So it makes
more sense to calculate variance improvemnet per numa node.
Signed-off-by: Cheng Li
Signed-off-by: Kevin Traynor
Co-authored-by: Kevin Traynor
---
Notes:
v2:
- Commit msg update
- Update doc for per numa variance
- Reword variance improment log msg
- Not break numa variance loop for debug purpose
Documentation/topics/dpdk/pmd.rst | 8 ++--
lib/dpif-netdev.c | 85 +++
2 files changed, 46 insertions(+), 47 deletions(-)
diff --git a/Documentation/topics/dpdk/pmd.rst
b/Documentation/topics/dpdk/pmd.rst
index b259cc8..c335757 100644
--- a/Documentation/topics/dpdk/pmd.rst
+++ b/Documentation/topics/dpdk/pmd.rst
@@ -278,10 +278,10 @@ If a PMD core is detected to be above the load threshold
and the minimum
pre-requisites are met, a dry-run using the current PMD assignment algorithm is
performed.
-The current variance of load between the PMD cores and estimated variance from
-the dry-run are both calculated. If the estimated dry-run variance is improved
-from the current one by the variance threshold, a new Rx queue to PMD
-assignment will be performed.
+For each numa node, the current variance of load between the PMD cores and
+estimated variance from the dry-run are both calculated. If any numa's
+estimated dry-run variance is improved from the current one by the variance
+threshold, a new Rx queue to PMD assignment will be performed.
For example, to set the variance improvement threshold to 40%::
diff --git a/lib/dpif-netdev.c b/lib/dpif-netdev.c
index 2c08a71..7ff923b 100644
--- a/lib/dpif-netdev.c
+++ b/lib/dpif-netdev.c
@@ -6076,39 +6076,33 @@ rxq_scheduling(struct dp_netdev *dp)
static uint64_t variance(uint64_t a[], int n);
static uint64_t
-sched_numa_list_variance(struct sched_numa_list *numa_list)
+sched_numa_variance(struct sched_numa *numa)
{
-struct sched_numa *numa;
uint64_t *percent_busy = NULL;
-unsigned total_pmds = 0;
int n_proc = 0;
uint64_t var;
-HMAP_FOR_EACH (numa, node, _list->numas) {
-total_pmds += numa->n_pmds;
-percent_busy = xrealloc(percent_busy,
-total_pmds * sizeof *percent_busy);
+percent_busy = xmalloc(numa->n_pmds * sizeof *percent_busy);
-for (unsigned i = 0; i < numa->n_pmds; i++) {
-struct sched_pmd *sched_pmd;
-uint64_t total_cycles = 0;
+for (unsigned i = 0; i < numa->n_pmds; i++) {
+struct sched_pmd *sched_pmd;
+uint64_t total_cycles = 0;
-sched_pmd = >pmds[i];
-/* Exclude isolated PMDs from variance calculations. */
-if (sched_pmd->isolated == true) {
-continue;
-}
-/* Get the total pmd cycles for an interval. */
-atomic_read_relaxed(_pmd->pmd->intrvl_cycles, _cycles);
+sched_pmd = >pmds[i];
+/* Exclude isolated PMDs from variance calculations. */
+if (sched_pmd->isolated == true) {
+continue;
+}
+/* Get the total pmd cycles for an interval. */
+atomic_read_relaxed(_pmd->pmd->intrvl_cycles, _cycles);
-if (total_cycles) {
-/* Estimate the cycles to cover all intervals. */
-total_cycles *= PMD_INTERVAL_MAX;
-percent_busy[n_proc++] = (sched_pmd->pmd_proc_cycles * 100)
- / total_cycles;
-} else {
-percent_busy[n_proc++] = 0;
-}
+if (total_cycles) {
+/* Estimate the cycles to cover all intervals. */
+total_cycles *= PMD_INTERVAL_MAX;
+percent_busy[n_proc++] = (sched_pmd->pmd_proc_cycles * 100)
+/ total_cycles;
+} else {
+percent_busy[n_proc++] = 0;
}
}
var = variance(percent_busy, n_proc);
@@ -6182,6 +6176,7 @@ pmd_rebalance_dry_run(struct dp_netdev *dp)
struct sched_numa_list
