[v2] examples/distributor: detect high frequency cores

Message ID 20190328131354.25222-1-david.hunt@intel.com
State Superseded, archived
Delegated to: Thomas Monjalon
Headers show
Series
  • [v2] examples/distributor: detect high frequency cores
Related show

Checks

Context Check Description
ci/intel-Performance-Testing success Performance Testing PASS
ci/mellanox-Performance-Testing fail Performance Testing issues
ci/Intel-compilation success Compilation OK
ci/checkpatch success coding style OK

Commit Message

David Hunt March 28, 2019, 1:13 p.m.
The distributor application is bottlenecked by the distributor core,
so if we can give more frequency to this core, then the overall
performance of the application may increase.

This patch uses the rte_power_get_capabilities() API to query the
cores provided in the core mask, and if any high frequency cores are
found (e.g. Turbo Boost is enabled), we will pin the distributor
workload to that core.

Signed-off-by: Liang Ma <liang.j.ma@intel.com>
Signed-off-by: David Hunt <david.hunt@intel.com>
---
 examples/distributor/main.c      | 204 ++++++++++++++++++++++++-------
 examples/distributor/meson.build |   2 +-
 2 files changed, 159 insertions(+), 47 deletions(-)

Comments

Burakov, Anatoly March 28, 2019, 1:58 p.m. | #1
On 28-Mar-19 1:13 PM, David Hunt wrote:
> The distributor application is bottlenecked by the distributor core,
> so if we can give more frequency to this core, then the overall
> performance of the application may increase.
> 
> This patch uses the rte_power_get_capabilities() API to query the
> cores provided in the core mask, and if any high frequency cores are
> found (e.g. Turbo Boost is enabled), we will pin the distributor
> workload to that core.
> 
> Signed-off-by: Liang Ma <liang.j.ma@intel.com>
> Signed-off-by: David Hunt <david.hunt@intel.com>
> ---

<...>

> +	if (power_lib_initialised)
> +		rte_power_exit(rte_lcore_id());
>   	printf("\nCore %u exiting tx task.\n", rte_lcore_id());
>   	return 0;
>   }
> @@ -575,9 +582,35 @@ lcore_worker(struct lcore_params *p)
>   		if (num > 0)
>   			app_stats.worker_bursts[p->worker_id][num-1]++;
>   	}
> +	if (power_lib_initialised)
> +		rte_power_exit(rte_lcore_id());
> +	rte_free(p);
>   	return 0;
>   }
>   
> +static int
> +init_power_library(void)
> +{
> +	int ret = 0, lcore_id;
> +	RTE_LCORE_FOREACH_SLAVE(lcore_id) {
> +		if (rte_lcore_is_enabled(lcore_id)) {

Please correct me if i'm wrong, but RTE_LCORE_FOREACH_SLAVE already 
checks if the lcore is enabled.

<...>

>   
> +	if (power_lib_initialised) {
> +		/*
> +		 * Here we'll pre-assign lcore ids to the rx, tx and
> +		 * distributor workloads if there's higher frequency
> +		 * on those cores e.g. if Turbo Boost is enabled.
> +		 * It's also worth mentioning that it will assign cores in a
> +		 * specific order, so that if there's less than three
> +		 * available, the higher frequency cores will go to the
> +		 * distributor first, then rx, then tx.
> +		 */
> +		RTE_LCORE_FOREACH_SLAVE(lcore_id) {
> +
> +			rte_power_get_capabilities(lcore_id, &lcore_cap);
> +
> +			if (lcore_cap.turbo == 1) {
> +				priority_num++;
> +				switch (priority_num) {
> +				case 1:
> +					distr_core_id = lcore_id;
> +					printf("Distributor on priority core %d\n",
> +							lcore_id);
> +					break;
> +				case 2:
> +					rx_core_id = lcore_id;
> +					printf("Rx on priority core %d\n",
> +							lcore_id);
> +					break;
> +				case 3:
> +					tx_core_id = lcore_id;
> +					printf("Tx on priority core %d\n",
> +							lcore_id);
> +					break;
> +				default:
> +					break;
> +				}

This seems to be doing the same thing as right below (assigning lcore 
id's in order), yet in one case you use a switch, and in the other you 
use a simple loop. I don't see priority_num used anywhere else, so you 
might as well simplify this loop to be similar to what you have below, 
with "skip-if-not-turbo, if not assigned, assign-and-continue" type flow.

Once that is fixed,

Reviewed-by: Anatoly Burakov <anatoly.burakov@intel.com>
David Hunt March 28, 2019, 2:42 p.m. | #2
On 28/3/2019 1:58 PM, Burakov, Anatoly wrote:
> On 28-Mar-19 1:13 PM, David Hunt wrote:
>> The distributor application is bottlenecked by the distributor core,
>> so if we can give more frequency to this core, then the overall
>> performance of the application may increase.
>>
>> This patch uses the rte_power_get_capabilities() API to query the
>> cores provided in the core mask, and if any high frequency cores are
>> found (e.g. Turbo Boost is enabled), we will pin the distributor
>> workload to that core.
>>
>> Signed-off-by: Liang Ma <liang.j.ma@intel.com>
>> Signed-off-by: David Hunt <david.hunt@intel.com>
>> ---
>
> <...>
>
>> +    if (power_lib_initialised)
>> +        rte_power_exit(rte_lcore_id());
>>       printf("\nCore %u exiting tx task.\n", rte_lcore_id());
>>       return 0;
>>   }
>> @@ -575,9 +582,35 @@ lcore_worker(struct lcore_params *p)
>>           if (num > 0)
>>               app_stats.worker_bursts[p->worker_id][num-1]++;
>>       }
>> +    if (power_lib_initialised)
>> +        rte_power_exit(rte_lcore_id());
>> +    rte_free(p);
>>       return 0;
>>   }
>>   +static int
>> +init_power_library(void)
>> +{
>> +    int ret = 0, lcore_id;
>> +    RTE_LCORE_FOREACH_SLAVE(lcore_id) {
>> +        if (rte_lcore_is_enabled(lcore_id)) {
>
> Please correct me if i'm wrong, but RTE_LCORE_FOREACH_SLAVE already 
> checks if the lcore is enabled.


You're correct, I'll fix in next version.


>
> <...>
>
>>   +    if (power_lib_initialised) {
>> +        /*
>> +         * Here we'll pre-assign lcore ids to the rx, tx and
>> +         * distributor workloads if there's higher frequency
>> +         * on those cores e.g. if Turbo Boost is enabled.
>> +         * It's also worth mentioning that it will assign cores in a
>> +         * specific order, so that if there's less than three
>> +         * available, the higher frequency cores will go to the
>> +         * distributor first, then rx, then tx.
>> +         */
>> +        RTE_LCORE_FOREACH_SLAVE(lcore_id) {
>> +
>> +            rte_power_get_capabilities(lcore_id, &lcore_cap);
>> +
>> +            if (lcore_cap.turbo == 1) {
>> +                priority_num++;
>> +                switch (priority_num) {
>> +                case 1:
>> +                    distr_core_id = lcore_id;
>> +                    printf("Distributor on priority core %d\n",
>> +                            lcore_id);
>> +                    break;
>> +                case 2:
>> +                    rx_core_id = lcore_id;
>> +                    printf("Rx on priority core %d\n",
>> +                            lcore_id);
>> +                    break;
>> +                case 3:
>> +                    tx_core_id = lcore_id;
>> +                    printf("Tx on priority core %d\n",
>> +                            lcore_id);
>> +                    break;
>> +                default:
>> +                    break;
>> +                }
>
> This seems to be doing the same thing as right below (assigning lcore 
> id's in order), yet in one case you use a switch, and in the other you 
> use a simple loop. I don't see priority_num used anywhere else, so you 
> might as well simplify this loop to be similar to what you have below, 
> with "skip-if-not-turbo, if not assigned, assign-and-continue" type flow.


There doing different things. The loop with the switch is looking for up 
to three priority cores, and storing those choices in distr_core_id, 
tx_core_id and rx_core_id. This is because we don't know which are the 
priority cores ahead of time. priority_num is used in the switch 
statement, and when it finds a priority core, it increments, so we know 
which variable to assign with the next available priority core. Imagine 
we have turbo enabled on cores 2,4 and 6. That's what I'm trying to 
solve here.

Then, when we get to the next loop, we're just assigning the 
non-priority cores if the three key workloads have not already been 
assigned a core, hence the simple loop, using the remaining cores.

I looked at simplifying the flow, but as far as I can see, I need two 
stages, a 'discovery' for the priority cores first, then whatever is 
left can be done in a normal loop.

Does that make sense, or my I missing an obvious refactor opportunity?


>
> Once that is fixed,
>
> Reviewed-by: Anatoly Burakov <anatoly.burakov@intel.com>
>
Burakov, Anatoly March 28, 2019, 3:10 p.m. | #3
On 28-Mar-19 2:42 PM, Hunt, David wrote:
> 
> On 28/3/2019 1:58 PM, Burakov, Anatoly wrote:
>> On 28-Mar-19 1:13 PM, David Hunt wrote:
>>> The distributor application is bottlenecked by the distributor core,
>>> so if we can give more frequency to this core, then the overall
>>> performance of the application may increase.
>>>
>>> This patch uses the rte_power_get_capabilities() API to query the
>>> cores provided in the core mask, and if any high frequency cores are
>>> found (e.g. Turbo Boost is enabled), we will pin the distributor
>>> workload to that core.
>>>
>>> Signed-off-by: Liang Ma <liang.j.ma@intel.com>
>>> Signed-off-by: David Hunt <david.hunt@intel.com>
>>> ---
>>
>> <...>
>>
>>> +    if (power_lib_initialised)
>>> +        rte_power_exit(rte_lcore_id());
>>>       printf("\nCore %u exiting tx task.\n", rte_lcore_id());
>>>       return 0;
>>>   }
>>> @@ -575,9 +582,35 @@ lcore_worker(struct lcore_params *p)
>>>           if (num > 0)
>>>               app_stats.worker_bursts[p->worker_id][num-1]++;
>>>       }
>>> +    if (power_lib_initialised)
>>> +        rte_power_exit(rte_lcore_id());
>>> +    rte_free(p);
>>>       return 0;
>>>   }
>>>   +static int
>>> +init_power_library(void)
>>> +{
>>> +    int ret = 0, lcore_id;
>>> +    RTE_LCORE_FOREACH_SLAVE(lcore_id) {
>>> +        if (rte_lcore_is_enabled(lcore_id)) {
>>
>> Please correct me if i'm wrong, but RTE_LCORE_FOREACH_SLAVE already 
>> checks if the lcore is enabled.
> 
> 
> You're correct, I'll fix in next version.
> 
> 
>>
>> <...>
>>
>>>   +    if (power_lib_initialised) {
>>> +        /*
>>> +         * Here we'll pre-assign lcore ids to the rx, tx and
>>> +         * distributor workloads if there's higher frequency
>>> +         * on those cores e.g. if Turbo Boost is enabled.
>>> +         * It's also worth mentioning that it will assign cores in a
>>> +         * specific order, so that if there's less than three
>>> +         * available, the higher frequency cores will go to the
>>> +         * distributor first, then rx, then tx.
>>> +         */
>>> +        RTE_LCORE_FOREACH_SLAVE(lcore_id) {
>>> +
>>> +            rte_power_get_capabilities(lcore_id, &lcore_cap);
>>> +
>>> +            if (lcore_cap.turbo == 1) {
>>> +                priority_num++;
>>> +                switch (priority_num) {
>>> +                case 1:
>>> +                    distr_core_id = lcore_id;
>>> +                    printf("Distributor on priority core %d\n",
>>> +                            lcore_id);
>>> +                    break;
>>> +                case 2:
>>> +                    rx_core_id = lcore_id;
>>> +                    printf("Rx on priority core %d\n",
>>> +                            lcore_id);
>>> +                    break;
>>> +                case 3:
>>> +                    tx_core_id = lcore_id;
>>> +                    printf("Tx on priority core %d\n",
>>> +                            lcore_id);
>>> +                    break;
>>> +                default:
>>> +                    break;
>>> +                }
>>
>> This seems to be doing the same thing as right below (assigning lcore 
>> id's in order), yet in one case you use a switch, and in the other you 
>> use a simple loop. I don't see priority_num used anywhere else, so you 
>> might as well simplify this loop to be similar to what you have below, 
>> with "skip-if-not-turbo, if not assigned, assign-and-continue" type flow.
> 
> 
> There doing different things. The loop with the switch is looking for up 
> to three priority cores, and storing those choices in distr_core_id, 
> tx_core_id and rx_core_id. This is because we don't know which are the 
> priority cores ahead of time. priority_num is used in the switch 
> statement, and when it finds a priority core, it increments, so we know 
> which variable to assign with the next available priority core. Imagine 
> we have turbo enabled on cores 2,4 and 6. That's what I'm trying to 
> solve here.
> 
> Then, when we get to the next loop, we're just assigning the 
> non-priority cores if the three key workloads have not already been 
> assigned a core, hence the simple loop, using the remaining cores.
> 
> I looked at simplifying the flow, but as far as I can see, I need two 
> stages, a 'discovery' for the priority cores first, then whatever is 
> left can be done in a normal loop.
> 
> Does that make sense, or my I missing an obvious refactor opportunity?

I don't see how this is different from what you're doing below.

You are looping over cores, checking if it's a priority core, and 
assigning any priority cores found to distributor, Rx and Tx cores, in 
that order.

Below, you're looping over cores, checking if the core is already 
assigned, and assigning these cores to distributor, Rx and Tx cores, in 
that order.

So, the only tangible difference between the two is 1) the check for 
whether the cores are already assigned (you don't need that because 
these cores *cannot* be attached - you haven't looped over them yet!), 
and 2) check for whether the core is priority.

Just to clarify: i'm not saying merge the two loops, that can't work :) 
I'm saying, drop the switch and rewrite it like this:

for (cores) {
	if (core not priority)
		continue;
	if (dist_core not assigned) {
		assign dist core;
		continue;
	}
	if (rx core not assigned) {
		assign rx core;
		continue;
	}
	if (tx core not assigned) {
		assign tx core;
		continue;
	}
}

The functionality would be equivalent to the current switch method, but 
the code would be much clearer :)

> 
> 
>>
>> Once that is fixed,
>>
>> Reviewed-by: Anatoly Burakov <anatoly.burakov@intel.com>
>>
>
David Hunt March 28, 2019, 3:20 p.m. | #4
On 28/3/2019 3:10 PM, Burakov, Anatoly wrote:
> On 28-Mar-19 2:42 PM, Hunt, David wrote:
>>
>> On 28/3/2019 1:58 PM, Burakov, Anatoly wrote:
>>> On 28-Mar-19 1:13 PM, David Hunt wrote:
>>>> The distributor application is bottlenecked by the distributor core,
>>>> so if we can give more frequency to this core, then the overall
>>>> performance of the application may increase.
>>>>
>>>> This patch uses the rte_power_get_capabilities() API to query the
>>>> cores provided in the core mask, and if any high frequency cores are
>>>> found (e.g. Turbo Boost is enabled), we will pin the distributor
>>>> workload to that core.
>>>>
>>>> Signed-off-by: Liang Ma <liang.j.ma@intel.com>
>>>> Signed-off-by: David Hunt <david.hunt@intel.com>
>>>> ---
>>>
>>> <...>
>>>
>>>> +    if (power_lib_initialised)
>>>> +        rte_power_exit(rte_lcore_id());
>>>>       printf("\nCore %u exiting tx task.\n", rte_lcore_id());
>>>>       return 0;
>>>>   }
>>>> @@ -575,9 +582,35 @@ lcore_worker(struct lcore_params *p)
>>>>           if (num > 0)
>>>> app_stats.worker_bursts[p->worker_id][num-1]++;
>>>>       }
>>>> +    if (power_lib_initialised)
>>>> +        rte_power_exit(rte_lcore_id());
>>>> +    rte_free(p);
>>>>       return 0;
>>>>   }
>>>>   +static int
>>>> +init_power_library(void)
>>>> +{
>>>> +    int ret = 0, lcore_id;
>>>> +    RTE_LCORE_FOREACH_SLAVE(lcore_id) {
>>>> +        if (rte_lcore_is_enabled(lcore_id)) {
>>>
>>> Please correct me if i'm wrong, but RTE_LCORE_FOREACH_SLAVE already 
>>> checks if the lcore is enabled.
>>
>>
>> You're correct, I'll fix in next version.
>>
>>
>>>
>>> <...>
>>>
>>>>   +    if (power_lib_initialised) {
>>>> +        /*
>>>> +         * Here we'll pre-assign lcore ids to the rx, tx and
>>>> +         * distributor workloads if there's higher frequency
>>>> +         * on those cores e.g. if Turbo Boost is enabled.
>>>> +         * It's also worth mentioning that it will assign cores in a
>>>> +         * specific order, so that if there's less than three
>>>> +         * available, the higher frequency cores will go to the
>>>> +         * distributor first, then rx, then tx.
>>>> +         */
>>>> +        RTE_LCORE_FOREACH_SLAVE(lcore_id) {
>>>> +
>>>> +            rte_power_get_capabilities(lcore_id, &lcore_cap);
>>>> +
>>>> +            if (lcore_cap.turbo == 1) {
>>>> +                priority_num++;
>>>> +                switch (priority_num) {
>>>> +                case 1:
>>>> +                    distr_core_id = lcore_id;
>>>> +                    printf("Distributor on priority core %d\n",
>>>> +                            lcore_id);
>>>> +                    break;
>>>> +                case 2:
>>>> +                    rx_core_id = lcore_id;
>>>> +                    printf("Rx on priority core %d\n",
>>>> +                            lcore_id);
>>>> +                    break;
>>>> +                case 3:
>>>> +                    tx_core_id = lcore_id;
>>>> +                    printf("Tx on priority core %d\n",
>>>> +                            lcore_id);
>>>> +                    break;
>>>> +                default:
>>>> +                    break;
>>>> +                }
>>>
>>> This seems to be doing the same thing as right below (assigning 
>>> lcore id's in order), yet in one case you use a switch, and in the 
>>> other you use a simple loop. I don't see priority_num used anywhere 
>>> else, so you might as well simplify this loop to be similar to what 
>>> you have below, with "skip-if-not-turbo, if not assigned, 
>>> assign-and-continue" type flow.
>>
>>
>> There doing different things. The loop with the switch is looking for 
>> up to three priority cores, and storing those choices in 
>> distr_core_id, tx_core_id and rx_core_id. This is because we don't 
>> know which are the priority cores ahead of time. priority_num is used 
>> in the switch statement, and when it finds a priority core, it 
>> increments, so we know which variable to assign with the next 
>> available priority core. Imagine we have turbo enabled on cores 2,4 
>> and 6. That's what I'm trying to solve here.
>>
>> Then, when we get to the next loop, we're just assigning the 
>> non-priority cores if the three key workloads have not already been 
>> assigned a core, hence the simple loop, using the remaining cores.
>>
>> I looked at simplifying the flow, but as far as I can see, I need two 
>> stages, a 'discovery' for the priority cores first, then whatever is 
>> left can be done in a normal loop.
>>
>> Does that make sense, or my I missing an obvious refactor opportunity?
>
> I don't see how this is different from what you're doing below.
>
> You are looping over cores, checking if it's a priority core, and 
> assigning any priority cores found to distributor, Rx and Tx cores, in 
> that order.
>
> Below, you're looping over cores, checking if the core is already 
> assigned, and assigning these cores to distributor, Rx and Tx cores, 
> in that order.
>
> So, the only tangible difference between the two is 1) the check for 
> whether the cores are already assigned (you don't need that because 
> these cores *cannot* be attached - you haven't looped over them yet!), 
> and 2) check for whether the core is priority.
>
> Just to clarify: i'm not saying merge the two loops, that can't work 
> :) I'm saying, drop the switch and rewrite it like this:
>

Ah, gocha now. Will do in next rev. Thanks!


> for (cores) {
>     if (core not priority)
>         continue;
>     if (dist_core not assigned) {
>         assign dist core;
>         continue;
>     }
>     if (rx core not assigned) {
>         assign rx core;
>         continue;
>     }
>     if (tx core not assigned) {
>         assign tx core;
>         continue;
>     }
> }
>
> The functionality would be equivalent to the current switch method, 
> but the code would be much clearer :)
>
>>
>>
>>>
>>> Once that is fixed,
>>>
>>> Reviewed-by: Anatoly Burakov <anatoly.burakov@intel.com>
>>>
>>
>
>

Patch

diff --git a/examples/distributor/main.c b/examples/distributor/main.c
index 03a05e3d9..1f8bcd672 100644
--- a/examples/distributor/main.c
+++ b/examples/distributor/main.c
@@ -16,6 +16,7 @@ 
 #include <rte_prefetch.h>
 #include <rte_distributor.h>
 #include <rte_pause.h>
+#include <rte_power.h>
 
 #define RX_RING_SIZE 1024
 #define TX_RING_SIZE 1024
@@ -37,6 +38,7 @@  volatile uint8_t quit_signal;
 volatile uint8_t quit_signal_rx;
 volatile uint8_t quit_signal_dist;
 volatile uint8_t quit_signal_work;
+unsigned int power_lib_initialised;
 
 static volatile struct app_stats {
 	struct {
@@ -281,6 +283,8 @@  lcore_rx(struct lcore_params *p)
 		if (++port == nb_ports)
 			port = 0;
 	}
+	if (power_lib_initialised)
+		rte_power_exit(rte_lcore_id());
 	/* set worker & tx threads quit flag */
 	printf("\nCore %u exiting rx task.\n", rte_lcore_id());
 	quit_signal = 1;
@@ -363,7 +367,8 @@  lcore_distributor(struct lcore_params *p)
 	}
 	printf("\nCore %u exiting distributor task.\n", rte_lcore_id());
 	quit_signal_work = 1;
-
+	if (power_lib_initialised)
+		rte_power_exit(rte_lcore_id());
 	rte_distributor_flush(d);
 	/* Unblock any returns so workers can exit */
 	rte_distributor_clear_returns(d);
@@ -435,6 +440,8 @@  lcore_tx(struct rte_ring *in_r)
 			}
 		}
 	}
+	if (power_lib_initialised)
+		rte_power_exit(rte_lcore_id());
 	printf("\nCore %u exiting tx task.\n", rte_lcore_id());
 	return 0;
 }
@@ -575,9 +582,35 @@  lcore_worker(struct lcore_params *p)
 		if (num > 0)
 			app_stats.worker_bursts[p->worker_id][num-1]++;
 	}
+	if (power_lib_initialised)
+		rte_power_exit(rte_lcore_id());
+	rte_free(p);
 	return 0;
 }
 
+static int
+init_power_library(void)
+{
+	int ret = 0, lcore_id;
+	RTE_LCORE_FOREACH_SLAVE(lcore_id) {
+		if (rte_lcore_is_enabled(lcore_id)) {
+			/* init power management library */
+			ret = rte_power_init(lcore_id);
+			if (ret) {
+				RTE_LOG(ERR, POWER,
+					"Library initialization failed on core %u\n",
+					lcore_id);
+				/*
+				 * Return on first failure, we'll fall back
+				 * to non-power operation
+				 */
+				return ret;
+			}
+		}
+	}
+	return ret;
+}
+
 /* display usage */
 static void
 print_usage(const char *prgname)
@@ -657,7 +690,9 @@  main(int argc, char *argv[])
 	struct rte_distributor *d;
 	struct rte_ring *dist_tx_ring;
 	struct rte_ring *rx_dist_ring;
-	unsigned lcore_id, worker_id = 0;
+	struct rte_power_core_capabilities lcore_cap;
+	unsigned int lcore_id, worker_id = 0, priority_num = 0;
+	int distr_core_id = -1, rx_core_id = -1, tx_core_id = -1;
 	unsigned nb_ports;
 	uint16_t portid;
 	uint16_t nb_ports_available;
@@ -687,6 +722,9 @@  main(int argc, char *argv[])
 				"1 lcore for packet TX\n"
 				"and at least 1 lcore for worker threads\n");
 
+	if (init_power_library() == 0)
+		power_lib_initialised = 1;
+
 	nb_ports = rte_eth_dev_count_avail();
 	if (nb_ports == 0)
 		rte_exit(EXIT_FAILURE, "Error: no ethernet ports detected\n");
@@ -742,54 +780,124 @@  main(int argc, char *argv[])
 	if (rx_dist_ring == NULL)
 		rte_exit(EXIT_FAILURE, "Cannot create output ring\n");
 
+	if (power_lib_initialised) {
+		/*
+		 * Here we'll pre-assign lcore ids to the rx, tx and
+		 * distributor workloads if there's higher frequency
+		 * on those cores e.g. if Turbo Boost is enabled.
+		 * It's also worth mentioning that it will assign cores in a
+		 * specific order, so that if there's less than three
+		 * available, the higher frequency cores will go to the
+		 * distributor first, then rx, then tx.
+		 */
+		RTE_LCORE_FOREACH_SLAVE(lcore_id) {
+
+			rte_power_get_capabilities(lcore_id, &lcore_cap);
+
+			if (lcore_cap.turbo == 1) {
+				priority_num++;
+				switch (priority_num) {
+				case 1:
+					distr_core_id = lcore_id;
+					printf("Distributor on priority core %d\n",
+							lcore_id);
+					break;
+				case 2:
+					rx_core_id = lcore_id;
+					printf("Rx on priority core %d\n",
+							lcore_id);
+					break;
+				case 3:
+					tx_core_id = lcore_id;
+					printf("Tx on priority core %d\n",
+							lcore_id);
+					break;
+				default:
+					break;
+				}
+			}
+		}
+	}
+
+	/*
+	 * If there's any of the key workloads left without an lcore_id
+	 * after the high performing core assignment above, pre-assign
+	 * them here.
+	 */
+	RTE_LCORE_FOREACH_SLAVE(lcore_id) {
+		if (lcore_id == (unsigned int)distr_core_id ||
+				lcore_id == (unsigned int)rx_core_id ||
+				lcore_id == (unsigned int)tx_core_id)
+			continue;
+		if (distr_core_id < 0) {
+			distr_core_id = lcore_id;
+			printf("Distributor on core %d\n", lcore_id);
+			continue;
+		}
+		if (rx_core_id < 0) {
+			rx_core_id = lcore_id;
+			printf("Rx on core %d\n", lcore_id);
+			continue;
+		}
+		if (tx_core_id < 0) {
+			tx_core_id = lcore_id;
+			printf("Tx on core %d\n", lcore_id);
+			continue;
+		}
+	}
+
+	printf(" tx id %d, dist id %d, rx id %d\n",
+			tx_core_id,
+			distr_core_id,
+			rx_core_id);
+
+	/*
+	 * Kick off all the worker threads first, avoiding the pre-assigned
+	 * lcore_ids for tx, rx and distributor workloads.
+	 */
 	RTE_LCORE_FOREACH_SLAVE(lcore_id) {
-		if (worker_id == rte_lcore_count() - 3) {
-			printf("Starting distributor on lcore_id %d\n",
-					lcore_id);
-			/* distributor core */
-			struct lcore_params *p =
-					rte_malloc(NULL, sizeof(*p), 0);
-			if (!p)
-				rte_panic("malloc failure\n");
-			*p = (struct lcore_params){worker_id, d,
-				rx_dist_ring, dist_tx_ring, mbuf_pool};
-			rte_eal_remote_launch(
-				(lcore_function_t *)lcore_distributor,
+		if (lcore_id == (unsigned int)distr_core_id ||
+				lcore_id == (unsigned int)rx_core_id ||
+				lcore_id == (unsigned int)tx_core_id)
+			continue;
+		printf("Starting thread %d as worker, lcore_id %d\n",
+				worker_id, lcore_id);
+		struct lcore_params *p =
+			rte_malloc(NULL, sizeof(*p), 0);
+		if (!p)
+			rte_panic("malloc failure\n");
+		*p = (struct lcore_params){worker_id++, d, rx_dist_ring,
+			dist_tx_ring, mbuf_pool};
+
+		rte_eal_remote_launch((lcore_function_t *)lcore_worker,
 				p, lcore_id);
-		} else if (worker_id == rte_lcore_count() - 4) {
-			printf("Starting tx  on worker_id %d, lcore_id %d\n",
-					worker_id, lcore_id);
-			/* tx core */
-			rte_eal_remote_launch((lcore_function_t *)lcore_tx,
-					dist_tx_ring, lcore_id);
-		} else if (worker_id == rte_lcore_count() - 2) {
-			printf("Starting rx on worker_id %d, lcore_id %d\n",
-					worker_id, lcore_id);
-			/* rx core */
-			struct lcore_params *p =
-					rte_malloc(NULL, sizeof(*p), 0);
-			if (!p)
-				rte_panic("malloc failure\n");
-			*p = (struct lcore_params){worker_id, d, rx_dist_ring,
-					dist_tx_ring, mbuf_pool};
-			rte_eal_remote_launch((lcore_function_t *)lcore_rx,
-					p, lcore_id);
-		} else {
-			printf("Starting worker on worker_id %d, lcore_id %d\n",
-					worker_id, lcore_id);
-			struct lcore_params *p =
-					rte_malloc(NULL, sizeof(*p), 0);
-			if (!p)
-				rte_panic("malloc failure\n");
-			*p = (struct lcore_params){worker_id, d, rx_dist_ring,
-					dist_tx_ring, mbuf_pool};
-
-			rte_eal_remote_launch((lcore_function_t *)lcore_worker,
-					p, lcore_id);
-		}
-		worker_id++;
 	}
 
+	/* Start tx core */
+	rte_eal_remote_launch((lcore_function_t *)lcore_tx,
+			dist_tx_ring, tx_core_id);
+
+	/* Start distributor core */
+	struct lcore_params *pd =
+		rte_malloc(NULL, sizeof(*pd), 0);
+	if (!pd)
+		rte_panic("malloc failure\n");
+	*pd = (struct lcore_params){worker_id++, d,
+		rx_dist_ring, dist_tx_ring, mbuf_pool};
+	rte_eal_remote_launch(
+			(lcore_function_t *)lcore_distributor,
+			pd, distr_core_id);
+
+	/* Start rx core */
+	struct lcore_params *pr =
+		rte_malloc(NULL, sizeof(*pr), 0);
+	if (!pr)
+		rte_panic("malloc failure\n");
+	*pr = (struct lcore_params){worker_id++, d, rx_dist_ring,
+		dist_tx_ring, mbuf_pool};
+	rte_eal_remote_launch((lcore_function_t *)lcore_rx,
+			pr, rx_core_id);
+
 	freq = rte_get_timer_hz();
 	t = rte_rdtsc() + freq;
 	while (!quit_signal_dist) {
@@ -806,5 +914,9 @@  main(int argc, char *argv[])
 	}
 
 	print_stats();
+
+	rte_free(pd);
+	rte_free(pr);
+
 	return 0;
 }
diff --git a/examples/distributor/meson.build b/examples/distributor/meson.build
index 88c001f56..8cf2ca1da 100644
--- a/examples/distributor/meson.build
+++ b/examples/distributor/meson.build
@@ -6,7 +6,7 @@ 
 # To build this example as a standalone application with an already-installed
 # DPDK instance, use 'make'
 
-deps += 'distributor'
+deps += ['distributor', 'power']
 sources = files(
 	'main.c'
 )