diff mbox

[ovs-dev,v6,4/6] dpif-netdev: Change rxq_scheduling to use rxq processing cycles.

Message ID 1503617885-32501-5-git-send-email-ktraynor@redhat.com
State Accepted
Delegated to: Darrell Ball
Headers show

Commit Message

Kevin Traynor Aug. 24, 2017, 11:38 p.m. UTC
Previously rxqs were assigned to pmds by round robin in
port/queue order.

Now that we have the processing cycles used for existing rxqs,
use that information to try and produced a better balanced
distribution of rxqs across pmds. i.e. given multiple pmds, the
rxqs which have consumed the largest amount of processing cycles
will be placed on different pmds.

The rxqs are sorted by their processing cycles and assigned (in
sorted order) round robin across pmds.

Signed-off-by: Kevin Traynor <ktraynor@redhat.com>
---
 Documentation/howto/dpdk.rst |   7 +++
 lib/dpif-netdev.c            | 123 ++++++++++++++++++++++++++++++++-----------
 2 files changed, 99 insertions(+), 31 deletions(-)

Comments

Ilya Maximets Aug. 28, 2017, 1:29 p.m. UTC | #1
> Previously rxqs were assigned to pmds by round robin in
> port/queue order.
> 
> Now that we have the processing cycles used for existing rxqs,
> use that information to try and produced a better balanced
> distribution of rxqs across pmds. i.e. given multiple pmds, the
> rxqs which have consumed the largest amount of processing cycles
> will be placed on different pmds.
> 
> The rxqs are sorted by their processing cycles and assigned (in
> sorted order) round robin across pmds.


One high level concern about this change:

Any change in any port or addition/deletion of PMD thread will likely
lead to rescheduling of all the ports to different PMD threads,
especially, if the load of threads is semi-equal. This will cause full
stop of packet processing for the time of reconfiguration and following
cache misses.

> 
> Signed-off-by: Kevin Traynor <ktraynor at redhat.com>
> ---
>  Documentation/howto/dpdk.rst |   7 +++
>  lib/dpif-netdev.c            | 123 ++++++++++++++++++++++++++++++++-----------
>  2 files changed, 99 insertions(+), 31 deletions(-)
> 
> diff --git a/Documentation/howto/dpdk.rst b/Documentation/howto/dpdk.rst
> index d7f6610..a67f3a1 100644
> --- a/Documentation/howto/dpdk.rst
> +++ b/Documentation/howto/dpdk.rst
> @@ -119,4 +119,11 @@ After that PMD threads on cores where RX queues was pinned will become
>    thread.
>  
> +If pmd-rxq-affinity is not set for rxqs, they will be assigned to pmds (cores)
> +automatically. The processing cycles that have been stored for each rxq
> +will be used where known to assign rxqs to pmd based on a round robin of the
> +sorted rxqs.
> +
> +Rxq to pmds assignment takes place whenever there are configuration changes.
> +
>  QoS
>  ---
> diff --git a/lib/dpif-netdev.c b/lib/dpif-netdev.c
> index cdf2662..4b1cb85 100644
> --- a/lib/dpif-netdev.c
> +++ b/lib/dpif-netdev.c
> @@ -715,4 +715,6 @@ static void
>  dp_netdev_rxq_set_intrvl_cycles(struct dp_netdev_rxq *rx,
>                             unsigned long long cycles);
> +static uint64_t
> +dp_netdev_rxq_get_intrvl_cycles(struct dp_netdev_rxq *rx, unsigned idx);
>  static void
>  dpif_netdev_xps_revalidate_pmd(const struct dp_netdev_pmd_thread *pmd,
> @@ -3169,4 +3171,12 @@ dp_netdev_rxq_set_intrvl_cycles(struct dp_netdev_rxq *rx,
>  }
>  
> +static uint64_t
> +dp_netdev_rxq_get_intrvl_cycles(struct dp_netdev_rxq *rx, unsigned idx)
> +{
> +    unsigned long long processing_cycles;
> +    atomic_read_relaxed(&rx->cycles_intrvl[idx], &processing_cycles);
> +    return processing_cycles;
> +}
> +
>  static int
>  dp_netdev_process_rxq_port(struct dp_netdev_pmd_thread *pmd,
> @@ -3355,8 +3365,37 @@ rr_numa_list_destroy(struct rr_numa_list *rr)
>  }
>  
> +/* Sort Rx Queues by the processing cycles they are consuming. */
> +static int
> +rxq_cycle_sort(const void *a, const void *b)
> +{
> +    struct dp_netdev_rxq * qa;
> +    struct dp_netdev_rxq * qb;

'*' should be placed near to variable.

> +    uint64_t total_qa, total_qb;
> +    unsigned i;
> +
> +    qa = *(struct dp_netdev_rxq **) a;
> +    qb = *(struct dp_netdev_rxq **) b;
> +
> +    total_qa = total_qb = 0;
> +    for (i = 0; i < PMD_RXQ_INTERVAL_MAX; i++) {
> +        total_qa += dp_netdev_rxq_get_intrvl_cycles(qa, i);
> +        total_qb += dp_netdev_rxq_get_intrvl_cycles(qb, i);

You're recalculating the sum of stats each time qsort calls comparison
function. This may lead to inconsistent results and, actually,
undefined result of sorting.

> +    }
> +    dp_netdev_rxq_set_cycles(qa, RXQ_CYCLES_PROC_HIST, total_qa);
> +    dp_netdev_rxq_set_cycles(qb, RXQ_CYCLES_PROC_HIST, total_qb);
> +
> +    if (total_qa >= total_qb) {
> +        return -1;
> +    }
> +    return 1;

Function should return '0' in case of equal values. Above code also
able to produce undefined sorting result.

> +}
> +
>  /* Assign pmds to queues.  If 'pinned' is true, assign pmds to pinned
>   * queues and marks the pmds as isolated.  Otherwise, assign non isolated
>   * pmds to unpinned queues.
>   *
> + * If 'pinned' is false queues will be sorted by processing cycles they are
> + * consuming and then assigned to pmds in round robin order.
> + *
>   * The function doesn't touch the pmd threads, it just stores the assignment
>   * in the 'pmd' member of each rxq. */
> @@ -3367,18 +3406,14 @@ rxq_scheduling(struct dp_netdev *dp, bool pinned) OVS_REQUIRES(dp->port_mutex)
>      struct rr_numa_list rr;
>      struct rr_numa *non_local_numa = NULL;
> -
> -    rr_numa_list_populate(dp, &rr);
> +    struct dp_netdev_rxq ** rxqs = NULL;
> +    int i, n_rxqs = 0;
> +    struct rr_numa *numa = NULL;
> +    int numa_id;

Can you merge some of above lines with the same type or at least
group them near each other.

>  
>      HMAP_FOR_EACH (port, node, &dp->ports) {
> -        struct rr_numa *numa;
> -        int numa_id;
> -
>          if (!netdev_is_pmd(port->netdev)) {
>              continue;
>          }
>  
> -        numa_id = netdev_get_numa_id(port->netdev);
> -        numa = rr_numa_list_lookup(&rr, numa_id);
> -
>          for (int qid = 0; qid < port->n_rxq; qid++) {
>              struct dp_netdev_rxq *q = &port->rxqs[qid];
> @@ -3398,34 +3433,60 @@ rxq_scheduling(struct dp_netdev *dp, bool pinned) OVS_REQUIRES(dp->port_mutex)
>                  }
>              } else if (!pinned && q->core_id == OVS_CORE_UNSPEC) {
> -                if (!numa) {
> -                    /* There are no pmds on the queue's local NUMA node.
> -                       Round-robin on the NUMA nodes that do have pmds. */
> -                    non_local_numa = rr_numa_list_next(&rr, non_local_numa);
> -                    if (!non_local_numa) {
> -                        VLOG_ERR("There is no available (non-isolated) pmd "
> -                                 "thread for port \'%s\' queue %d. This queue "
> -                                 "will not be polled. Is pmd-cpu-mask set to "
> -                                 "zero? Or are all PMDs isolated to other "
> -                                 "queues?", netdev_get_name(port->netdev),
> -                                 qid);
> -                        continue;
> -                    }
> -                    q->pmd = rr_numa_get_pmd(non_local_numa);
> -                    VLOG_WARN("There's no available (non-isolated) pmd thread "
> -                              "on numa node %d. Queue %d on port \'%s\' will "
> -                              "be assigned to the pmd on core %d "
> -                              "(numa node %d). Expect reduced performance.",
> -                              numa_id, qid, netdev_get_name(port->netdev),
> -                              q->pmd->core_id, q->pmd->numa_id);
> +                if (n_rxqs == 0) {
> +                    rxqs = xmalloc(sizeof *rxqs);
>                  } else {
> -                    /* Assign queue to the next (round-robin) PMD on it's local
> -                       NUMA node. */
> -                    q->pmd = rr_numa_get_pmd(numa);
> +                    rxqs = xrealloc(rxqs, sizeof *rxqs * (n_rxqs + 1));
>                  }
> +                /* Store the queue. */
> +                rxqs[n_rxqs++] = q;
>              }
>          }
>      }
>  
> +    if (n_rxqs > 1) {
> +        /* Sort the queues in order of the processing cycles
> +         * they consumed during their last pmd interval. */
> +        qsort(rxqs, n_rxqs, sizeof *rxqs, rxq_cycle_sort);
> +    }
> +
> +    rr_numa_list_populate(dp, &rr);
> +    /* Assign the sorted queues to pmds in round robin. */
> +    for (i = 0; i < n_rxqs; i++) {
> +        numa_id = netdev_get_numa_id(rxqs[i]->port->netdev);
> +        numa = rr_numa_list_lookup(&rr, numa_id);
> +        if (!numa) {
> +            /* There are no pmds on the queue's local NUMA node.
> +               Round robin on the NUMA nodes that do have pmds. */
> +            non_local_numa = rr_numa_list_next(&rr, non_local_numa);
> +            if (!non_local_numa) {
> +                VLOG_ERR("There is no available (non-isolated) pmd "
> +                         "thread for port \'%s\' queue %d. This queue "
> +                         "will not be polled. Is pmd-cpu-mask set to "
> +                         "zero? Or are all PMDs isolated to other "
> +                         "queues?", netdev_rxq_get_name(rxqs[i]->rx),
> +                         netdev_rxq_get_queue_id(rxqs[i]->rx));
> +                continue;
> +            }
> +            rxqs[i]->pmd = rr_numa_get_pmd(non_local_numa);
> +            VLOG_WARN("There's no available (non-isolated) pmd thread "
> +                      "on numa node %d. Queue %d on port \'%s\' will "
> +                      "be assigned to the pmd on core %d "
> +                      "(numa node %d). Expect reduced performance.",
> +                      numa_id, netdev_rxq_get_queue_id(rxqs[i]->rx),
> +                      netdev_rxq_get_name(rxqs[i]->rx),
> +                      rxqs[i]->pmd->core_id, rxqs[i]->pmd->numa_id);
> +        } else {
> +        rxqs[i]->pmd = rr_numa_get_pmd(numa);
> +        VLOG_INFO("Core %d on numa node %d assigned port \'%s\' "
> +                  "rx queue %d (measured processing cycles %"PRIu64").",
> +                  rxqs[i]->pmd->core_id, numa_id,
> +                  netdev_rxq_get_name(rxqs[i]->rx),
> +                  netdev_rxq_get_queue_id(rxqs[i]->rx),
> +                  dp_netdev_rxq_get_cycles(rxqs[i], RXQ_CYCLES_PROC_HIST));
> +        }
> +    }
> +
>      rr_numa_list_destroy(&rr);
> +    free(rxqs);
>  }
>  
> -- 
> 1.8.3.
Kevin Traynor Aug. 30, 2017, 4:56 p.m. UTC | #2
On 08/28/2017 02:29 PM, Ilya Maximets wrote:
>> Previously rxqs were assigned to pmds by round robin in
>> port/queue order.
>>
>> Now that we have the processing cycles used for existing rxqs,
>> use that information to try and produced a better balanced
>> distribution of rxqs across pmds. i.e. given multiple pmds, the
>> rxqs which have consumed the largest amount of processing cycles
>> will be placed on different pmds.
>>
>> The rxqs are sorted by their processing cycles and assigned (in
>> sorted order) round robin across pmds.
> 
> 
> One high level concern about this change:
> 
> Any change in any port or addition/deletion of PMD thread will likely
> lead to rescheduling of all the ports to different PMD threads,
> especially, if the load of threads is semi-equal. This will cause full
> stop of packet processing for the time of reconfiguration and following
> cache misses.
> 

This could happen with port changes or addition/deletion of pmds in the
previous scheme too. I agree it could be more likely to happen in some
circumstances with the new scheme, but dropping packets in the steady
state operation because of an imbalanced system without any way to
address it is a bigger problem IMHO. I guess it may be possible in the
future to refine and factor in existing assignments to some extent.

>>
>> Signed-off-by: Kevin Traynor <ktraynor at redhat.com>
>> ---
>>  Documentation/howto/dpdk.rst |   7 +++
>>  lib/dpif-netdev.c            | 123 ++++++++++++++++++++++++++++++++-----------
>>  2 files changed, 99 insertions(+), 31 deletions(-)
>>
>> diff --git a/Documentation/howto/dpdk.rst b/Documentation/howto/dpdk.rst
>> index d7f6610..a67f3a1 100644
>> --- a/Documentation/howto/dpdk.rst
>> +++ b/Documentation/howto/dpdk.rst
>> @@ -119,4 +119,11 @@ After that PMD threads on cores where RX queues was pinned will become
>>    thread.
>>  
>> +If pmd-rxq-affinity is not set for rxqs, they will be assigned to pmds (cores)
>> +automatically. The processing cycles that have been stored for each rxq
>> +will be used where known to assign rxqs to pmd based on a round robin of the
>> +sorted rxqs.
>> +
>> +Rxq to pmds assignment takes place whenever there are configuration changes.
>> +
>>  QoS
>>  ---
>> diff --git a/lib/dpif-netdev.c b/lib/dpif-netdev.c
>> index cdf2662..4b1cb85 100644
>> --- a/lib/dpif-netdev.c
>> +++ b/lib/dpif-netdev.c
>> @@ -715,4 +715,6 @@ static void
>>  dp_netdev_rxq_set_intrvl_cycles(struct dp_netdev_rxq *rx,
>>                             unsigned long long cycles);
>> +static uint64_t
>> +dp_netdev_rxq_get_intrvl_cycles(struct dp_netdev_rxq *rx, unsigned idx);
>>  static void
>>  dpif_netdev_xps_revalidate_pmd(const struct dp_netdev_pmd_thread *pmd,
>> @@ -3169,4 +3171,12 @@ dp_netdev_rxq_set_intrvl_cycles(struct dp_netdev_rxq *rx,
>>  }
>>  
>> +static uint64_t
>> +dp_netdev_rxq_get_intrvl_cycles(struct dp_netdev_rxq *rx, unsigned idx)
>> +{
>> +    unsigned long long processing_cycles;
>> +    atomic_read_relaxed(&rx->cycles_intrvl[idx], &processing_cycles);
>> +    return processing_cycles;
>> +}
>> +
>>  static int
>>  dp_netdev_process_rxq_port(struct dp_netdev_pmd_thread *pmd,
>> @@ -3355,8 +3365,37 @@ rr_numa_list_destroy(struct rr_numa_list *rr)
>>  }
>>  
>> +/* Sort Rx Queues by the processing cycles they are consuming. */
>> +static int
>> +rxq_cycle_sort(const void *a, const void *b)
>> +{
>> +    struct dp_netdev_rxq * qa;
>> +    struct dp_netdev_rxq * qb;
> 
> '*' should be placed near to variable.
> 

sure, I will fix this.

>> +    uint64_t total_qa, total_qb;
>> +    unsigned i;
>> +
>> +    qa = *(struct dp_netdev_rxq **) a;
>> +    qb = *(struct dp_netdev_rxq **) b;
>> +
>> +    total_qa = total_qb = 0;
>> +    for (i = 0; i < PMD_RXQ_INTERVAL_MAX; i++) {
>> +        total_qa += dp_netdev_rxq_get_intrvl_cycles(qa, i);
>> +        total_qb += dp_netdev_rxq_get_intrvl_cycles(qb, i);
> 
> You're recalculating the sum of stats each time qsort calls comparison
> function. This may lead to inconsistent results and, actually,
> undefined result of sorting.
> 

My thinking was it is a corner case and is ok because even if the sort
was a bit inconsistent, the data used for each comparison would be more
recent so there would be a better overall result. But I take the point
that maybe it's better to have a data set that doesn't change. I will
send a patch to calculate total cycles for each queue before calling qsort.

>> +    }
>> +    dp_netdev_rxq_set_cycles(qa, RXQ_CYCLES_PROC_HIST, total_qa);
>> +    dp_netdev_rxq_set_cycles(qb, RXQ_CYCLES_PROC_HIST, total_qb);
>> +
>> +    if (total_qa >= total_qb) {
>> +        return -1;
>> +    }
>> +    return 1;
> 
> Function should return '0' in case of equal values. Above code also
> able to produce undefined sorting result.
> 

I will add this in a new patch for rxq_cycle_sort.

>> +}
>> +
>>  /* Assign pmds to queues.  If 'pinned' is true, assign pmds to pinned
>>   * queues and marks the pmds as isolated.  Otherwise, assign non isolated
>>   * pmds to unpinned queues.
>>   *
>> + * If 'pinned' is false queues will be sorted by processing cycles they are
>> + * consuming and then assigned to pmds in round robin order.
>> + *
>>   * The function doesn't touch the pmd threads, it just stores the assignment
>>   * in the 'pmd' member of each rxq. */
>> @@ -3367,18 +3406,14 @@ rxq_scheduling(struct dp_netdev *dp, bool pinned) OVS_REQUIRES(dp->port_mutex)
>>      struct rr_numa_list rr;
>>      struct rr_numa *non_local_numa = NULL;
>> -
>> -    rr_numa_list_populate(dp, &rr);
>> +    struct dp_netdev_rxq ** rxqs = NULL;
>> +    int i, n_rxqs = 0;
>> +    struct rr_numa *numa = NULL;
>> +    int numa_id;
> 
> Can you merge some of above lines with the same type or at least
> group them near each other.
> 

I think it's ok as is.

>>  
>>      HMAP_FOR_EACH (port, node, &dp->ports) {
>> -        struct rr_numa *numa;
>> -        int numa_id;
>> -
>>          if (!netdev_is_pmd(port->netdev)) {
>>              continue;
>>          }
>>  
>> -        numa_id = netdev_get_numa_id(port->netdev);
>> -        numa = rr_numa_list_lookup(&rr, numa_id);
>> -
>>          for (int qid = 0; qid < port->n_rxq; qid++) {
>>              struct dp_netdev_rxq *q = &port->rxqs[qid];
>> @@ -3398,34 +3433,60 @@ rxq_scheduling(struct dp_netdev *dp, bool pinned) OVS_REQUIRES(dp->port_mutex)
>>                  }
>>              } else if (!pinned && q->core_id == OVS_CORE_UNSPEC) {
>> -                if (!numa) {
>> -                    /* There are no pmds on the queue's local NUMA node.
>> -                       Round-robin on the NUMA nodes that do have pmds. */
>> -                    non_local_numa = rr_numa_list_next(&rr, non_local_numa);
>> -                    if (!non_local_numa) {
>> -                        VLOG_ERR("There is no available (non-isolated) pmd "
>> -                                 "thread for port \'%s\' queue %d. This queue "
>> -                                 "will not be polled. Is pmd-cpu-mask set to "
>> -                                 "zero? Or are all PMDs isolated to other "
>> -                                 "queues?", netdev_get_name(port->netdev),
>> -                                 qid);
>> -                        continue;
>> -                    }
>> -                    q->pmd = rr_numa_get_pmd(non_local_numa);
>> -                    VLOG_WARN("There's no available (non-isolated) pmd thread "
>> -                              "on numa node %d. Queue %d on port \'%s\' will "
>> -                              "be assigned to the pmd on core %d "
>> -                              "(numa node %d). Expect reduced performance.",
>> -                              numa_id, qid, netdev_get_name(port->netdev),
>> -                              q->pmd->core_id, q->pmd->numa_id);
>> +                if (n_rxqs == 0) {
>> +                    rxqs = xmalloc(sizeof *rxqs);
>>                  } else {
>> -                    /* Assign queue to the next (round-robin) PMD on it's local
>> -                       NUMA node. */
>> -                    q->pmd = rr_numa_get_pmd(numa);
>> +                    rxqs = xrealloc(rxqs, sizeof *rxqs * (n_rxqs + 1));
>>                  }
>> +                /* Store the queue. */
>> +                rxqs[n_rxqs++] = q;
>>              }
>>          }
>>      }
>>  
>> +    if (n_rxqs > 1) {
>> +        /* Sort the queues in order of the processing cycles
>> +         * they consumed during their last pmd interval. */
>> +        qsort(rxqs, n_rxqs, sizeof *rxqs, rxq_cycle_sort);
>> +    }
>> +
>> +    rr_numa_list_populate(dp, &rr);
>> +    /* Assign the sorted queues to pmds in round robin. */
>> +    for (i = 0; i < n_rxqs; i++) {
>> +        numa_id = netdev_get_numa_id(rxqs[i]->port->netdev);
>> +        numa = rr_numa_list_lookup(&rr, numa_id);
>> +        if (!numa) {
>> +            /* There are no pmds on the queue's local NUMA node.
>> +               Round robin on the NUMA nodes that do have pmds. */
>> +            non_local_numa = rr_numa_list_next(&rr, non_local_numa);
>> +            if (!non_local_numa) {
>> +                VLOG_ERR("There is no available (non-isolated) pmd "
>> +                         "thread for port \'%s\' queue %d. This queue "
>> +                         "will not be polled. Is pmd-cpu-mask set to "
>> +                         "zero? Or are all PMDs isolated to other "
>> +                         "queues?", netdev_rxq_get_name(rxqs[i]->rx),
>> +                         netdev_rxq_get_queue_id(rxqs[i]->rx));
>> +                continue;
>> +            }
>> +            rxqs[i]->pmd = rr_numa_get_pmd(non_local_numa);
>> +            VLOG_WARN("There's no available (non-isolated) pmd thread "
>> +                      "on numa node %d. Queue %d on port \'%s\' will "
>> +                      "be assigned to the pmd on core %d "
>> +                      "(numa node %d). Expect reduced performance.",
>> +                      numa_id, netdev_rxq_get_queue_id(rxqs[i]->rx),
>> +                      netdev_rxq_get_name(rxqs[i]->rx),
>> +                      rxqs[i]->pmd->core_id, rxqs[i]->pmd->numa_id);
>> +        } else {
>> +        rxqs[i]->pmd = rr_numa_get_pmd(numa);
>> +        VLOG_INFO("Core %d on numa node %d assigned port \'%s\' "
>> +                  "rx queue %d (measured processing cycles %"PRIu64").",
>> +                  rxqs[i]->pmd->core_id, numa_id,
>> +                  netdev_rxq_get_name(rxqs[i]->rx),
>> +                  netdev_rxq_get_queue_id(rxqs[i]->rx),
>> +                  dp_netdev_rxq_get_cycles(rxqs[i], RXQ_CYCLES_PROC_HIST));
>> +        }
>> +    }
>> +
>>      rr_numa_list_destroy(&rr);
>> +    free(rxqs);
>>  }
>>  
>> -- 
>> 1.8.3.
diff mbox

Patch

diff --git a/Documentation/howto/dpdk.rst b/Documentation/howto/dpdk.rst
index d7f6610..a67f3a1 100644
--- a/Documentation/howto/dpdk.rst
+++ b/Documentation/howto/dpdk.rst
@@ -119,4 +119,11 @@  After that PMD threads on cores where RX queues was pinned will become
   thread.
 
+If pmd-rxq-affinity is not set for rxqs, they will be assigned to pmds (cores)
+automatically. The processing cycles that have been stored for each rxq
+will be used where known to assign rxqs to pmd based on a round robin of the
+sorted rxqs.
+
+Rxq to pmds assignment takes place whenever there are configuration changes.
+
 QoS
 ---
diff --git a/lib/dpif-netdev.c b/lib/dpif-netdev.c
index cdf2662..4b1cb85 100644
--- a/lib/dpif-netdev.c
+++ b/lib/dpif-netdev.c
@@ -715,4 +715,6 @@  static void
 dp_netdev_rxq_set_intrvl_cycles(struct dp_netdev_rxq *rx,
                            unsigned long long cycles);
+static uint64_t
+dp_netdev_rxq_get_intrvl_cycles(struct dp_netdev_rxq *rx, unsigned idx);
 static void
 dpif_netdev_xps_revalidate_pmd(const struct dp_netdev_pmd_thread *pmd,
@@ -3169,4 +3171,12 @@  dp_netdev_rxq_set_intrvl_cycles(struct dp_netdev_rxq *rx,
 }
 
+static uint64_t
+dp_netdev_rxq_get_intrvl_cycles(struct dp_netdev_rxq *rx, unsigned idx)
+{
+    unsigned long long processing_cycles;
+    atomic_read_relaxed(&rx->cycles_intrvl[idx], &processing_cycles);
+    return processing_cycles;
+}
+
 static int
 dp_netdev_process_rxq_port(struct dp_netdev_pmd_thread *pmd,
@@ -3355,8 +3365,37 @@  rr_numa_list_destroy(struct rr_numa_list *rr)
 }
 
+/* Sort Rx Queues by the processing cycles they are consuming. */
+static int
+rxq_cycle_sort(const void *a, const void *b)
+{
+    struct dp_netdev_rxq * qa;
+    struct dp_netdev_rxq * qb;
+    uint64_t total_qa, total_qb;
+    unsigned i;
+
+    qa = *(struct dp_netdev_rxq **) a;
+    qb = *(struct dp_netdev_rxq **) b;
+
+    total_qa = total_qb = 0;
+    for (i = 0; i < PMD_RXQ_INTERVAL_MAX; i++) {
+        total_qa += dp_netdev_rxq_get_intrvl_cycles(qa, i);
+        total_qb += dp_netdev_rxq_get_intrvl_cycles(qb, i);
+    }
+    dp_netdev_rxq_set_cycles(qa, RXQ_CYCLES_PROC_HIST, total_qa);
+    dp_netdev_rxq_set_cycles(qb, RXQ_CYCLES_PROC_HIST, total_qb);
+
+    if (total_qa >= total_qb) {
+        return -1;
+    }
+    return 1;
+}
+
 /* Assign pmds to queues.  If 'pinned' is true, assign pmds to pinned
  * queues and marks the pmds as isolated.  Otherwise, assign non isolated
  * pmds to unpinned queues.
  *
+ * If 'pinned' is false queues will be sorted by processing cycles they are
+ * consuming and then assigned to pmds in round robin order.
+ *
  * The function doesn't touch the pmd threads, it just stores the assignment
  * in the 'pmd' member of each rxq. */
@@ -3367,18 +3406,14 @@  rxq_scheduling(struct dp_netdev *dp, bool pinned) OVS_REQUIRES(dp->port_mutex)
     struct rr_numa_list rr;
     struct rr_numa *non_local_numa = NULL;
-
-    rr_numa_list_populate(dp, &rr);
+    struct dp_netdev_rxq ** rxqs = NULL;
+    int i, n_rxqs = 0;
+    struct rr_numa *numa = NULL;
+    int numa_id;
 
     HMAP_FOR_EACH (port, node, &dp->ports) {
-        struct rr_numa *numa;
-        int numa_id;
-
         if (!netdev_is_pmd(port->netdev)) {
             continue;
         }
 
-        numa_id = netdev_get_numa_id(port->netdev);
-        numa = rr_numa_list_lookup(&rr, numa_id);
-
         for (int qid = 0; qid < port->n_rxq; qid++) {
             struct dp_netdev_rxq *q = &port->rxqs[qid];
@@ -3398,34 +3433,60 @@  rxq_scheduling(struct dp_netdev *dp, bool pinned) OVS_REQUIRES(dp->port_mutex)
                 }
             } else if (!pinned && q->core_id == OVS_CORE_UNSPEC) {
-                if (!numa) {
-                    /* There are no pmds on the queue's local NUMA node.
-                       Round-robin on the NUMA nodes that do have pmds. */
-                    non_local_numa = rr_numa_list_next(&rr, non_local_numa);
-                    if (!non_local_numa) {
-                        VLOG_ERR("There is no available (non-isolated) pmd "
-                                 "thread for port \'%s\' queue %d. This queue "
-                                 "will not be polled. Is pmd-cpu-mask set to "
-                                 "zero? Or are all PMDs isolated to other "
-                                 "queues?", netdev_get_name(port->netdev),
-                                 qid);
-                        continue;
-                    }
-                    q->pmd = rr_numa_get_pmd(non_local_numa);
-                    VLOG_WARN("There's no available (non-isolated) pmd thread "
-                              "on numa node %d. Queue %d on port \'%s\' will "
-                              "be assigned to the pmd on core %d "
-                              "(numa node %d). Expect reduced performance.",
-                              numa_id, qid, netdev_get_name(port->netdev),
-                              q->pmd->core_id, q->pmd->numa_id);
+                if (n_rxqs == 0) {
+                    rxqs = xmalloc(sizeof *rxqs);
                 } else {
-                    /* Assign queue to the next (round-robin) PMD on it's local
-                       NUMA node. */
-                    q->pmd = rr_numa_get_pmd(numa);
+                    rxqs = xrealloc(rxqs, sizeof *rxqs * (n_rxqs + 1));
                 }
+                /* Store the queue. */
+                rxqs[n_rxqs++] = q;
             }
         }
     }
 
+    if (n_rxqs > 1) {
+        /* Sort the queues in order of the processing cycles
+         * they consumed during their last pmd interval. */
+        qsort(rxqs, n_rxqs, sizeof *rxqs, rxq_cycle_sort);
+    }
+
+    rr_numa_list_populate(dp, &rr);
+    /* Assign the sorted queues to pmds in round robin. */
+    for (i = 0; i < n_rxqs; i++) {
+        numa_id = netdev_get_numa_id(rxqs[i]->port->netdev);
+        numa = rr_numa_list_lookup(&rr, numa_id);
+        if (!numa) {
+            /* There are no pmds on the queue's local NUMA node.
+               Round robin on the NUMA nodes that do have pmds. */
+            non_local_numa = rr_numa_list_next(&rr, non_local_numa);
+            if (!non_local_numa) {
+                VLOG_ERR("There is no available (non-isolated) pmd "
+                         "thread for port \'%s\' queue %d. This queue "
+                         "will not be polled. Is pmd-cpu-mask set to "
+                         "zero? Or are all PMDs isolated to other "
+                         "queues?", netdev_rxq_get_name(rxqs[i]->rx),
+                         netdev_rxq_get_queue_id(rxqs[i]->rx));
+                continue;
+            }
+            rxqs[i]->pmd = rr_numa_get_pmd(non_local_numa);
+            VLOG_WARN("There's no available (non-isolated) pmd thread "
+                      "on numa node %d. Queue %d on port \'%s\' will "
+                      "be assigned to the pmd on core %d "
+                      "(numa node %d). Expect reduced performance.",
+                      numa_id, netdev_rxq_get_queue_id(rxqs[i]->rx),
+                      netdev_rxq_get_name(rxqs[i]->rx),
+                      rxqs[i]->pmd->core_id, rxqs[i]->pmd->numa_id);
+        } else {
+        rxqs[i]->pmd = rr_numa_get_pmd(numa);
+        VLOG_INFO("Core %d on numa node %d assigned port \'%s\' "
+                  "rx queue %d (measured processing cycles %"PRIu64").",
+                  rxqs[i]->pmd->core_id, numa_id,
+                  netdev_rxq_get_name(rxqs[i]->rx),
+                  netdev_rxq_get_queue_id(rxqs[i]->rx),
+                  dp_netdev_rxq_get_cycles(rxqs[i], RXQ_CYCLES_PROC_HIST));
+        }
+    }
+
     rr_numa_list_destroy(&rr);
+    free(rxqs);
 }