power: add traffic pattern aware power control
authorLiang Ma <liang.j.ma@intel.com>
Fri, 19 Oct 2018 11:07:18 +0000 (12:07 +0100)
committerThomas Monjalon <thomas@monjalon.net>
Thu, 25 Oct 2018 23:55:07 +0000 (01:55 +0200)
1. Abstract

For packet processing workloads such as DPDK polling is continuous.
This means CPU cores always show 100% busy independent of how much work
those cores are doing. It is critical to accurately determine how busy
a core is hugely important for the following reasons:

   * No indication of overload conditions.

   * User does not know how much real load is on a system, resulting
     in wasted energy as no power management is utilized.

Compared to the original l3fwd-power design, instead of going to sleep
after detecting an empty poll, the new mechanism just lowers the core
frequency. As a result, the application does not stop polling the device,
which leads to improved handling of bursts of traffic.

When the system become busy, the empty poll mechanism can also increase the
core frequency (including turbo) to do best effort for intensive traffic.
This gives us more flexible and balanced traffic awareness over the
standard l3fwd-power application.

2. Proposed solution

The proposed solution focuses on how many times empty polls are executed.
The less the number of empty polls, means current core is busy with
processing workload, therefore, the higher frequency is needed. The high
empty poll number indicates the current core not doing any real work
therefore, we can lower the frequency to safe power.

In the current implementation, each core has 1 empty-poll counter which
assume 1 core is dedicated to 1 queue. This will need to be expanded in the
future to support multiple queues per core.

2.1 Power state definition:

LOW:  Not currently used, reserved for future use.

MED:  the frequency is used to process modest traffic workload.

HIGH: the frequency is used to process busy traffic workload.

2.2 There are two phases to establish the power management system:

a.Initialization/Training phase. The training phase is necessary
  in order to figure out the system polling baseline numbers from
  idle to busy. The highest poll count will be during idle, where
  all polls are empty. These poll counts will be different between
  systems due to the many possible processor micro-arch, cache
  and device configurations, hence the training phase.
  In the training phase, traffic is blocked so the training
  algorithm can average the empty-poll numbers for the LOW, MED and
  HIGH  power states in order to create a baseline.
  The core's counter are collected every 10ms, and the Training
  phase will take 2 seconds.
  Training is disabled as default configuration. The default
  parameter is applied. Sample App still can trigger training
  if that's needed. Once the training phase has been executed once on
  a system, the application can then be started with the relevant
  thresholds provided on the command line, allowing the application
  to start passing start traffic immediately

b.Normal phase. Traffic starts immediately based on the default
  thresholds, or based on the user supplied thresholds via the
  command line parameters. The run-time poll counts are compared with
  the baseline and the decision will be taken to move to MED power
  state or HIGH power state. The counters are calculated every 10ms.

3. Proposed  API

1.  rte_power_empty_poll_stat_init(struct ep_params **eptr,
uint8_t *freq_tlb, struct ep_policy *policy);
which is used to initialize the power management system.
 
2.  rte_power_empty_poll_stat_free(void);
which is used to free the resource hold by power management system.
 
3.  rte_power_empty_poll_stat_update(unsigned int lcore_id);
which is used to update specific core empty poll counter, not thread safe
 
4.  rte_power_poll_stat_update(unsigned int lcore_id, uint8_t nb_pkt);
which is used to update specific core valid poll counter, not thread safe
 
5.  rte_power_empty_poll_stat_fetch(unsigned int lcore_id);
which is used to get specific core empty poll counter.
 
6.  rte_power_poll_stat_fetch(unsigned int lcore_id);
which is used to get specific core valid poll counter.

7.  rte_empty_poll_detection(struct rte_timer *tim, void *arg);
which is used to detect empty poll state changes then take action.

Signed-off-by: Liang Ma <liang.j.ma@intel.com>
Reviewed-by: Lei Yao <lei.a.yao@intel.com>
Acked-by: David Hunt <david.hunt@intel.com>
doc/guides/prog_guide/power_man.rst
doc/guides/rel_notes/release_18_11.rst
lib/librte_power/Makefile
lib/librte_power/meson.build
lib/librte_power/rte_power_empty_poll.c [new file with mode: 0644]
lib/librte_power/rte_power_empty_poll.h [new file with mode: 0644]
lib/librte_power/rte_power_version.map

index eba1cc6..68b7e8b 100644 (file)
@@ -106,6 +106,92 @@ User Cases
 
 The power management mechanism is used to save power when performing L3 forwarding.
 
+
+Empty Poll API
+--------------
+
+Abstract
+~~~~~~~~
+
+For packet processing workloads such as DPDK polling is continuous.
+This means CPU cores always show 100% busy independent of how much work
+those cores are doing. It is critical to accurately determine how busy
+a core is hugely important for the following reasons:
+
+        * No indication of overload conditions
+        * User does not know how much real load is on a system, resulting
+          in wasted energy as no power management is utilized
+
+Compared to the original l3fwd-power design, instead of going to sleep
+after detecting an empty poll, the new mechanism just lowers the core frequency.
+As a result, the application does not stop polling the device, which leads
+to improved handling of bursts of traffic.
+
+When the system become busy, the empty poll mechanism can also increase the core
+frequency (including turbo) to do best effort for intensive traffic. This gives
+us more flexible and balanced traffic awareness over the standard l3fwd-power
+application.
+
+
+Proposed Solution
+~~~~~~~~~~~~~~~~~
+The proposed solution focuses on how many times empty polls are executed.
+The less the number of empty polls, means current core is busy with processing
+workload, therefore, the higher frequency is needed. The high empty poll number
+indicates the current core not doing any real work therefore, we can lower the
+frequency to safe power.
+
+In the current implementation, each core has 1 empty-poll counter which assume
+1 core is dedicated to 1 queue. This will need to be expanded in the future to
+support multiple queues per core.
+
+Power state definition:
+^^^^^^^^^^^^^^^^^^^^^^^
+
+* LOW:  Not currently used, reserved for future use.
+
+* MED:  the frequency is used to process modest traffic workload.
+
+* HIGH: the frequency is used to process busy traffic workload.
+
+There are two phases to establish the power management system:
+^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+* Training phase. This phase is used to measure the optimal frequency
+  change thresholds for a given system. The thresholds will differ from
+  system to system due to differences in processor micro-architecture,
+  cache and device configurations.
+  In this phase, the user must ensure that no traffic can enter the
+  system so that counts can be measured for empty polls at low, medium
+  and high frequencies. Each frequency is measured for two seconds.
+  Once the training phase is complete, the threshold numbers are
+  displayed, and normal mode resumes, and traffic can be allowed into
+  the system. These threshold number can be used on the command line
+  when starting the application in normal mode to avoid re-training
+  every time.
+
+* Normal phase. Every 10ms the run-time counters are compared
+  to the supplied threshold values, and the decision will be made
+  whether to move to a different power state (by adjusting the
+  frequency).
+
+API Overview for Empty Poll Power Management
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+* **State Init**: initialize the power management system.
+
+* **State Free**: free the resource hold by power management system.
+
+* **Update Empty Poll Counter**: update the empty poll counter.
+
+* **Update Valid Poll Counter**: update the valid poll counter.
+
+* **Set the Fequence Index**: update the power state/frequency mapping.
+
+* **Detect empty poll state change**: empty poll state change detection algorithm then take action.
+
+User Cases
+----------
+The mechanism can applied to any device which is based on polling. e.g. NIC, FPGA.
+
 References
 ----------
 
index 0a90144..988cfb3 100644 (file)
@@ -190,6 +190,16 @@ New Features
   collisions.  In addition, the internal hashing algorithm was changed to use
   partial-key hashing to improve memory efficiency and lookup performance.
 
+* **Added Traffic Pattern Aware Power Control Library**
+
+  Added an experimental library. This extend Power Library and provide
+  empty_poll APIs. This feature measure how many times empty_poll are
+  executed per core, use the number of empty polls as a hint for system
+  power management.
+
+  See the :doc:`../prog_guide/power_man` section of the DPDK Programmers
+  Guide document for more information.
+
 * **Added ability to switch queue deferred start flag on testpmd app.**
 
   Added a console command to testpmd app, giving ability to switch
index 6f85e88..9bec668 100644 (file)
@@ -7,7 +7,7 @@ include $(RTE_SDK)/mk/rte.vars.mk
 LIB = librte_power.a
 
 CFLAGS += $(WERROR_FLAGS) -I$(SRCDIR) -O3 -fno-strict-aliasing
-LDLIBS += -lrte_eal
+LDLIBS += -lrte_eal -lrte_timer
 
 EXPORT_MAP := rte_power_version.map
 
@@ -16,8 +16,9 @@ LIBABIVER := 1
 # all source are stored in SRCS-y
 SRCS-$(CONFIG_RTE_LIBRTE_POWER) := rte_power.c power_acpi_cpufreq.c
 SRCS-$(CONFIG_RTE_LIBRTE_POWER) += power_kvm_vm.c guest_channel.c
+SRCS-$(CONFIG_RTE_LIBRTE_POWER) += rte_power_empty_poll.c
 
 # install this header file
-SYMLINK-$(CONFIG_RTE_LIBRTE_POWER)-include := rte_power.h
+SYMLINK-$(CONFIG_RTE_LIBRTE_POWER)-include := rte_power.h  rte_power_empty_poll.h
 
 include $(RTE_SDK)/mk/rte.lib.mk
index 253173f..9ed8b56 100644 (file)
@@ -5,5 +5,7 @@ if host_machine.system() != 'linux'
        build = false
 endif
 sources = files('rte_power.c', 'power_acpi_cpufreq.c',
-               'power_kvm_vm.c', 'guest_channel.c')
-headers = files('rte_power.h')
+               'power_kvm_vm.c', 'guest_channel.c',
+               'rte_power_empty_poll.c')
+headers = files('rte_power.h','rte_power_empty_poll.h')
+deps += ['timer']
diff --git a/lib/librte_power/rte_power_empty_poll.c b/lib/librte_power/rte_power_empty_poll.c
new file mode 100644 (file)
index 0000000..c1e10e0
--- /dev/null
@@ -0,0 +1,544 @@
+/* SPDX-License-Identifier: BSD-3-Clause
+ * Copyright(c) 2010-2018 Intel Corporation
+ */
+
+#include <string.h>
+
+#include <rte_lcore.h>
+#include <rte_cycles.h>
+#include <rte_atomic.h>
+#include <rte_malloc.h>
+
+#include "rte_power.h"
+#include "rte_power_empty_poll.h"
+
+#define INTERVALS_PER_SECOND 100     /* (10ms) */
+#define SECONDS_TO_TRAIN_FOR 2
+#define DEFAULT_MED_TO_HIGH_PERCENT_THRESHOLD 70
+#define DEFAULT_HIGH_TO_MED_PERCENT_THRESHOLD 30
+#define DEFAULT_CYCLES_PER_PACKET 800
+
+static struct ep_params *ep_params;
+static uint32_t med_to_high_threshold = DEFAULT_MED_TO_HIGH_PERCENT_THRESHOLD;
+static uint32_t high_to_med_threshold = DEFAULT_HIGH_TO_MED_PERCENT_THRESHOLD;
+
+static uint32_t avail_freqs[RTE_MAX_LCORE][NUM_FREQS];
+
+static uint32_t total_avail_freqs[RTE_MAX_LCORE];
+
+static uint32_t freq_index[NUM_FREQ];
+
+static uint32_t
+get_freq_index(enum freq_val index)
+{
+       return freq_index[index];
+}
+
+
+static int
+set_power_freq(int lcore_id, enum freq_val freq, bool specific_freq)
+{
+       int err = 0;
+       uint32_t power_freq_index;
+       if (!specific_freq)
+               power_freq_index = get_freq_index(freq);
+       else
+               power_freq_index = freq;
+
+       err = rte_power_set_freq(lcore_id, power_freq_index);
+
+       return err;
+}
+
+
+static inline void __attribute__((always_inline))
+exit_training_state(struct priority_worker *poll_stats)
+{
+       RTE_SET_USED(poll_stats);
+}
+
+static inline void __attribute__((always_inline))
+enter_training_state(struct priority_worker *poll_stats)
+{
+       poll_stats->iter_counter = 0;
+       poll_stats->cur_freq = LOW;
+       poll_stats->queue_state = TRAINING;
+}
+
+static inline void __attribute__((always_inline))
+enter_normal_state(struct priority_worker *poll_stats)
+{
+       /* Clear the averages arrays and strs */
+       memset(poll_stats->edpi_av, 0, sizeof(poll_stats->edpi_av));
+       poll_stats->ec = 0;
+       memset(poll_stats->ppi_av, 0, sizeof(poll_stats->ppi_av));
+       poll_stats->pc = 0;
+
+       poll_stats->cur_freq = MED;
+       poll_stats->iter_counter = 0;
+       poll_stats->threshold_ctr = 0;
+       poll_stats->queue_state = MED_NORMAL;
+       RTE_LOG(INFO, POWER, "Set the power freq to MED\n");
+       set_power_freq(poll_stats->lcore_id, MED, false);
+
+       poll_stats->thresh[MED].threshold_percent = med_to_high_threshold;
+       poll_stats->thresh[HGH].threshold_percent = high_to_med_threshold;
+}
+
+static inline void __attribute__((always_inline))
+enter_busy_state(struct priority_worker *poll_stats)
+{
+       memset(poll_stats->edpi_av, 0, sizeof(poll_stats->edpi_av));
+       poll_stats->ec = 0;
+       memset(poll_stats->ppi_av, 0, sizeof(poll_stats->ppi_av));
+       poll_stats->pc = 0;
+
+       poll_stats->cur_freq = HGH;
+       poll_stats->iter_counter = 0;
+       poll_stats->threshold_ctr = 0;
+       poll_stats->queue_state = HGH_BUSY;
+       set_power_freq(poll_stats->lcore_id, HGH, false);
+}
+
+static inline void __attribute__((always_inline))
+enter_purge_state(struct priority_worker *poll_stats)
+{
+       poll_stats->iter_counter = 0;
+       poll_stats->queue_state = LOW_PURGE;
+}
+
+static inline void __attribute__((always_inline))
+set_state(struct priority_worker *poll_stats,
+               enum queue_state new_state)
+{
+       enum queue_state old_state = poll_stats->queue_state;
+       if (old_state != new_state) {
+
+               /* Call any old state exit functions */
+               if (old_state == TRAINING)
+                       exit_training_state(poll_stats);
+
+               /* Call any new state entry functions */
+               if (new_state == TRAINING)
+                       enter_training_state(poll_stats);
+               if (new_state == MED_NORMAL)
+                       enter_normal_state(poll_stats);
+               if (new_state == HGH_BUSY)
+                       enter_busy_state(poll_stats);
+               if (new_state == LOW_PURGE)
+                       enter_purge_state(poll_stats);
+       }
+}
+
+static inline void __attribute__((always_inline))
+set_policy(struct priority_worker *poll_stats,
+               struct ep_policy *policy)
+{
+       set_state(poll_stats, policy->state);
+
+       if (policy->state == TRAINING)
+               return;
+
+       poll_stats->thresh[MED_NORMAL].base_edpi = policy->med_base_edpi;
+       poll_stats->thresh[HGH_BUSY].base_edpi = policy->hgh_base_edpi;
+
+       poll_stats->thresh[MED_NORMAL].trained = true;
+       poll_stats->thresh[HGH_BUSY].trained = true;
+
+}
+
+static void
+update_training_stats(struct priority_worker *poll_stats,
+               uint32_t freq,
+               bool specific_freq,
+               uint32_t max_train_iter)
+{
+       RTE_SET_USED(specific_freq);
+
+       char pfi_str[32];
+       uint64_t p0_empty_deq;
+
+       sprintf(pfi_str, "%02d", freq);
+
+       if (poll_stats->cur_freq == freq &&
+                       poll_stats->thresh[freq].trained == false) {
+               if (poll_stats->thresh[freq].cur_train_iter == 0) {
+
+                       set_power_freq(poll_stats->lcore_id,
+                                       freq, specific_freq);
+
+                       poll_stats->empty_dequeues_prev =
+                               poll_stats->empty_dequeues;
+
+                       poll_stats->thresh[freq].cur_train_iter++;
+
+                       return;
+               } else if (poll_stats->thresh[freq].cur_train_iter
+                               <= max_train_iter) {
+
+                       p0_empty_deq = poll_stats->empty_dequeues -
+                               poll_stats->empty_dequeues_prev;
+
+                       poll_stats->empty_dequeues_prev =
+                               poll_stats->empty_dequeues;
+
+                       poll_stats->thresh[freq].base_edpi += p0_empty_deq;
+                       poll_stats->thresh[freq].cur_train_iter++;
+
+               } else {
+                       if (poll_stats->thresh[freq].trained == false) {
+                               poll_stats->thresh[freq].base_edpi =
+                                       poll_stats->thresh[freq].base_edpi /
+                                       max_train_iter;
+
+                               /* Add on a factor of 0.05%
+                                * this should remove any
+                                * false negatives when the system is 0% busy
+                                */
+                               poll_stats->thresh[freq].base_edpi +=
+                               poll_stats->thresh[freq].base_edpi / 2000;
+
+                               poll_stats->thresh[freq].trained = true;
+                               poll_stats->cur_freq++;
+
+                       }
+               }
+       }
+}
+
+static inline uint32_t __attribute__((always_inline))
+update_stats(struct priority_worker *poll_stats)
+{
+       uint64_t tot_edpi = 0, tot_ppi = 0;
+       uint32_t j, percent;
+
+       struct priority_worker *s = poll_stats;
+
+       uint64_t cur_edpi = s->empty_dequeues - s->empty_dequeues_prev;
+
+       s->empty_dequeues_prev = s->empty_dequeues;
+
+       uint64_t ppi = s->num_dequeue_pkts - s->num_dequeue_pkts_prev;
+
+       s->num_dequeue_pkts_prev = s->num_dequeue_pkts;
+
+       if (s->thresh[s->cur_freq].base_edpi < cur_edpi) {
+
+               /* edpi mean empty poll counter difference per interval */
+               RTE_LOG(DEBUG, POWER, "cur_edpi is too large "
+                               "cur edpi %ld "
+                               "base edpi %ld\n",
+                               cur_edpi,
+                               s->thresh[s->cur_freq].base_edpi);
+               /* Value to make us fail need debug log*/
+               return 1000UL;
+       }
+
+       s->edpi_av[s->ec++ % BINS_AV] = cur_edpi;
+       s->ppi_av[s->pc++ % BINS_AV] = ppi;
+
+       for (j = 0; j < BINS_AV; j++) {
+               tot_edpi += s->edpi_av[j];
+               tot_ppi += s->ppi_av[j];
+       }
+
+       tot_edpi = tot_edpi / BINS_AV;
+
+       percent = 100 - (uint32_t)(((float)tot_edpi /
+                       (float)s->thresh[s->cur_freq].base_edpi) * 100);
+
+       return (uint32_t)percent;
+}
+
+
+static inline void  __attribute__((always_inline))
+update_stats_normal(struct priority_worker *poll_stats)
+{
+       uint32_t percent;
+
+       if (poll_stats->thresh[poll_stats->cur_freq].base_edpi == 0) {
+
+               enum freq_val cur_freq = poll_stats->cur_freq;
+
+               /* edpi mean empty poll counter difference per interval */
+               RTE_LOG(DEBUG, POWER, "cure freq is %d, edpi is %lu\n",
+                               cur_freq,
+                               poll_stats->thresh[cur_freq].base_edpi);
+               return;
+       }
+
+       percent = update_stats(poll_stats);
+
+       if (percent > 100) {
+               /* edpi mean empty poll counter difference per interval */
+               RTE_LOG(DEBUG, POWER, "Edpi is bigger than threshold\n");
+               return;
+       }
+
+       if (poll_stats->cur_freq == LOW)
+               RTE_LOG(INFO, POWER, "Purge Mode is not currently supported\n");
+       else if (poll_stats->cur_freq == MED) {
+
+               if (percent >
+                       poll_stats->thresh[MED].threshold_percent) {
+
+                       if (poll_stats->threshold_ctr < INTERVALS_PER_SECOND)
+                               poll_stats->threshold_ctr++;
+                       else {
+                               set_state(poll_stats, HGH_BUSY);
+                               RTE_LOG(INFO, POWER, "MOVE to HGH\n");
+                       }
+
+               } else {
+                       /* reset */
+                       poll_stats->threshold_ctr = 0;
+               }
+
+       } else if (poll_stats->cur_freq == HGH) {
+
+               if (percent <
+                               poll_stats->thresh[HGH].threshold_percent) {
+
+                       if (poll_stats->threshold_ctr < INTERVALS_PER_SECOND)
+                               poll_stats->threshold_ctr++;
+                       else {
+                               set_state(poll_stats, MED_NORMAL);
+                               RTE_LOG(INFO, POWER, "MOVE to MED\n");
+                       }
+               } else {
+                       /* reset */
+                       poll_stats->threshold_ctr = 0;
+               }
+
+       }
+}
+
+static int
+empty_poll_training(struct priority_worker *poll_stats,
+               uint32_t max_train_iter)
+{
+
+       if (poll_stats->iter_counter < INTERVALS_PER_SECOND) {
+               poll_stats->iter_counter++;
+               return 0;
+       }
+
+
+       update_training_stats(poll_stats,
+                       LOW,
+                       false,
+                       max_train_iter);
+
+       update_training_stats(poll_stats,
+                       MED,
+                       false,
+                       max_train_iter);
+
+       update_training_stats(poll_stats,
+                       HGH,
+                       false,
+                       max_train_iter);
+
+
+       if (poll_stats->thresh[LOW].trained == true
+                       && poll_stats->thresh[MED].trained == true
+                       && poll_stats->thresh[HGH].trained == true) {
+
+               set_state(poll_stats, MED_NORMAL);
+
+               RTE_LOG(INFO, POWER, "LOW threshold is %lu\n",
+                               poll_stats->thresh[LOW].base_edpi);
+
+               RTE_LOG(INFO, POWER, "MED threshold is %lu\n",
+                               poll_stats->thresh[MED].base_edpi);
+
+
+               RTE_LOG(INFO, POWER, "HIGH threshold is %lu\n",
+                               poll_stats->thresh[HGH].base_edpi);
+
+               RTE_LOG(INFO, POWER, "Training is Complete for %d\n",
+                               poll_stats->lcore_id);
+       }
+
+       return 0;
+}
+
+void __rte_experimental
+rte_empty_poll_detection(struct rte_timer *tim, void *arg)
+{
+
+       uint32_t i;
+
+       struct priority_worker *poll_stats;
+
+       RTE_SET_USED(tim);
+
+       RTE_SET_USED(arg);
+
+       for (i = 0; i < NUM_NODES; i++) {
+
+               poll_stats = &(ep_params->wrk_data.wrk_stats[i]);
+
+               if (rte_lcore_is_enabled(poll_stats->lcore_id) == 0)
+                       continue;
+
+               switch (poll_stats->queue_state) {
+               case(TRAINING):
+                       empty_poll_training(poll_stats,
+                                       ep_params->max_train_iter);
+                       break;
+
+               case(HGH_BUSY):
+               case(MED_NORMAL):
+                       update_stats_normal(poll_stats);
+                       break;
+
+               case(LOW_PURGE):
+                       break;
+               default:
+                       break;
+
+               }
+
+       }
+
+}
+
+int __rte_experimental
+rte_power_empty_poll_stat_init(struct ep_params **eptr, uint8_t *freq_tlb,
+               struct ep_policy *policy)
+{
+       uint32_t i;
+       /* Allocate the ep_params structure */
+       ep_params = rte_zmalloc_socket(NULL,
+                       sizeof(struct ep_params),
+                       0,
+                       rte_socket_id());
+
+       if (!ep_params)
+               return -1;
+
+       if (freq_tlb == NULL) {
+               freq_index[LOW] = 14;
+               freq_index[MED] = 9;
+               freq_index[HGH] = 1;
+       } else {
+               freq_index[LOW] = freq_tlb[LOW];
+               freq_index[MED] = freq_tlb[MED];
+               freq_index[HGH] = freq_tlb[HGH];
+       }
+
+       RTE_LOG(INFO, POWER, "Initialize the Empty Poll\n");
+
+       /* Train for pre-defined period */
+       ep_params->max_train_iter = INTERVALS_PER_SECOND * SECONDS_TO_TRAIN_FOR;
+
+       struct stats_data *w = &ep_params->wrk_data;
+
+       *eptr = ep_params;
+
+       /* initialize all wrk_stats state */
+       for (i = 0; i < NUM_NODES; i++) {
+
+               if (rte_lcore_is_enabled(i) == 0)
+                       continue;
+               /*init the freqs table */
+               total_avail_freqs[i] = rte_power_freqs(i,
+                               avail_freqs[i],
+                               NUM_FREQS);
+
+               RTE_LOG(INFO, POWER, "total avail freq is %d , lcoreid %d\n",
+                               total_avail_freqs[i],
+                               i);
+
+               if (get_freq_index(LOW) > total_avail_freqs[i])
+                       return -1;
+
+               if (rte_get_master_lcore() != i) {
+                       w->wrk_stats[i].lcore_id = i;
+                       set_policy(&w->wrk_stats[i], policy);
+               }
+       }
+
+       return 0;
+}
+
+void __rte_experimental
+rte_power_empty_poll_stat_free(void)
+{
+
+       RTE_LOG(INFO, POWER, "Close the Empty Poll\n");
+
+       if (ep_params != NULL)
+               rte_free(ep_params);
+}
+
+int __rte_experimental
+rte_power_empty_poll_stat_update(unsigned int lcore_id)
+{
+       struct priority_worker *poll_stats;
+
+       if (lcore_id >= NUM_NODES)
+               return -1;
+
+       poll_stats = &(ep_params->wrk_data.wrk_stats[lcore_id]);
+
+       if (poll_stats->lcore_id == 0)
+               poll_stats->lcore_id = lcore_id;
+
+       poll_stats->empty_dequeues++;
+
+       return 0;
+}
+
+int __rte_experimental
+rte_power_poll_stat_update(unsigned int lcore_id, uint8_t nb_pkt)
+{
+
+       struct priority_worker *poll_stats;
+
+       if (lcore_id >= NUM_NODES)
+               return -1;
+
+       poll_stats = &(ep_params->wrk_data.wrk_stats[lcore_id]);
+
+       if (poll_stats->lcore_id == 0)
+               poll_stats->lcore_id = lcore_id;
+
+       poll_stats->num_dequeue_pkts += nb_pkt;
+
+       return 0;
+}
+
+
+uint64_t __rte_experimental
+rte_power_empty_poll_stat_fetch(unsigned int lcore_id)
+{
+       struct priority_worker *poll_stats;
+
+       if (lcore_id >= NUM_NODES)
+               return -1;
+
+       poll_stats = &(ep_params->wrk_data.wrk_stats[lcore_id]);
+
+       if (poll_stats->lcore_id == 0)
+               poll_stats->lcore_id = lcore_id;
+
+       return poll_stats->empty_dequeues;
+}
+
+uint64_t __rte_experimental
+rte_power_poll_stat_fetch(unsigned int lcore_id)
+{
+       struct priority_worker *poll_stats;
+
+       if (lcore_id >= NUM_NODES)
+               return -1;
+
+       poll_stats = &(ep_params->wrk_data.wrk_stats[lcore_id]);
+
+       if (poll_stats->lcore_id == 0)
+               poll_stats->lcore_id = lcore_id;
+
+       return poll_stats->num_dequeue_pkts;
+}
diff --git a/lib/librte_power/rte_power_empty_poll.h b/lib/librte_power/rte_power_empty_poll.h
new file mode 100644 (file)
index 0000000..c1ad5c2
--- /dev/null
@@ -0,0 +1,223 @@
+/* SPDX-License-Identifier: BSD-3-Clause
+ * Copyright(c) 2010-2018 Intel Corporation
+ */
+
+#ifndef _RTE_EMPTY_POLL_H
+#define _RTE_EMPTY_POLL_H
+
+/**
+ * @file
+ * RTE Power Management
+ */
+#include <stdint.h>
+#include <stdbool.h>
+
+#include <rte_common.h>
+#include <rte_byteorder.h>
+#include <rte_log.h>
+#include <rte_string_fns.h>
+#include <rte_power.h>
+#include <rte_timer.h>
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#define NUM_FREQS  RTE_MAX_LCORE_FREQS
+
+#define BINS_AV 4 /* Has to be ^2 */
+
+#define DROP (NUM_DIRECTIONS * NUM_DEVICES)
+
+#define NUM_PRIORITIES          2
+
+#define NUM_NODES         256  /* Max core number*/
+
+/* Processor Power State */
+enum freq_val {
+       LOW,
+       MED,
+       HGH,
+       NUM_FREQ = NUM_FREQS
+};
+
+
+/* Queue Polling State */
+enum queue_state {
+       TRAINING, /* NO TRAFFIC */
+       MED_NORMAL,   /* MED */
+       HGH_BUSY,     /* HIGH */
+       LOW_PURGE,    /* LOW */
+};
+
+/* Queue Stats */
+struct freq_threshold {
+
+       uint64_t base_edpi;
+       bool trained;
+       uint32_t threshold_percent;
+       uint32_t cur_train_iter;
+};
+
+/* Each Worder Thread Empty Poll Stats */
+struct priority_worker {
+
+       /* Current dequeue and throughput counts */
+       /* These 2 are written to by the worker threads */
+       /* So keep them on their own cache line */
+       uint64_t empty_dequeues;
+       uint64_t num_dequeue_pkts;
+
+       enum queue_state queue_state;
+
+       uint64_t empty_dequeues_prev;
+       uint64_t num_dequeue_pkts_prev;
+
+       /* Used for training only */
+       struct freq_threshold thresh[NUM_FREQ];
+       enum freq_val cur_freq;
+
+       /* bucket arrays to calculate the averages */
+       /* edpi mean empty poll counter difference per interval */
+       uint64_t edpi_av[BINS_AV];
+       /* empty poll counter */
+       uint32_t ec;
+       /* ppi mean valid poll counter per interval */
+       uint64_t ppi_av[BINS_AV];
+       /* valid poll counter */
+       uint32_t pc;
+
+       uint32_t lcore_id;
+       uint32_t iter_counter;
+       uint32_t threshold_ctr;
+       uint32_t display_ctr;
+       uint8_t  dev_id;
+
+} __rte_cache_aligned;
+
+
+struct stats_data {
+
+       struct priority_worker wrk_stats[NUM_NODES];
+
+       /* flag to stop rx threads processing packets until training over */
+       bool start_rx;
+
+};
+
+/* Empty Poll Parameters */
+struct ep_params {
+
+       /* Timer related stuff */
+       uint64_t interval_ticks;
+       uint32_t max_train_iter;
+
+       struct rte_timer timer0;
+       struct stats_data wrk_data;
+};
+
+
+/* Sample App Init information */
+struct ep_policy {
+
+       uint64_t med_base_edpi;
+       uint64_t hgh_base_edpi;
+
+       enum queue_state state;
+};
+
+
+
+/**
+ * Initialize the power management system.
+ *
+ * @param eptr
+ *   the structure of empty poll configuration
+ * @param freq_tlb
+ *   the power state/frequency mapping table
+ * @param policy
+ *   the initialization policy from sample app
+ *
+ * @return
+ *  - 0 on success.
+ *  - Negative on error.
+ */
+int __rte_experimental
+rte_power_empty_poll_stat_init(struct ep_params **eptr, uint8_t *freq_tlb,
+               struct ep_policy *policy);
+
+/**
+ * Free the resource hold by power management system.
+ */
+void __rte_experimental
+rte_power_empty_poll_stat_free(void);
+
+/**
+ * Update specific core empty poll counter
+ * It's not thread safe.
+ *
+ * @param lcore_id
+ *  lcore id
+ *
+ * @return
+ *  - 0 on success.
+ *  - Negative on error.
+ */
+int __rte_experimental
+rte_power_empty_poll_stat_update(unsigned int lcore_id);
+
+/**
+ * Update specific core valid poll counter, not thread safe.
+ *
+ * @param lcore_id
+ *  lcore id.
+ * @param nb_pkt
+ *  The packet number of one valid poll.
+ *
+ * @return
+ *  - 0 on success.
+ *  - Negative on error.
+ */
+int __rte_experimental
+rte_power_poll_stat_update(unsigned int lcore_id, uint8_t nb_pkt);
+
+/**
+ * Fetch specific core empty poll counter.
+ *
+ * @param lcore_id
+ *  lcore id
+ *
+ * @return
+ *  Current lcore empty poll counter value.
+ */
+uint64_t __rte_experimental
+rte_power_empty_poll_stat_fetch(unsigned int lcore_id);
+
+/**
+ * Fetch specific core valid poll counter.
+ *
+ * @param lcore_id
+ *  lcore id
+ *
+ * @return
+ *  Current lcore valid poll counter value.
+ */
+uint64_t __rte_experimental
+rte_power_poll_stat_fetch(unsigned int lcore_id);
+
+/**
+ * Empty poll  state change detection function
+ *
+ * @param  tim
+ *  The timer structure
+ * @param  arg
+ *  The customized parameter
+ */
+void  __rte_experimental
+rte_empty_poll_detection(struct rte_timer *tim, void *arg);
+
+#ifdef __cplusplus
+}
+#endif
+
+#endif
index dd587df..17a083b 100644 (file)
@@ -33,3 +33,16 @@ DPDK_18.08 {
        rte_power_get_capabilities;
 
 } DPDK_17.11;
+
+EXPERIMENTAL {
+        global:
+
+        rte_empty_poll_detection;
+        rte_power_empty_poll_stat_fetch;
+        rte_power_empty_poll_stat_free;
+        rte_power_empty_poll_stat_init;
+        rte_power_empty_poll_stat_update;
+        rte_power_poll_stat_fetch;
+        rte_power_poll_stat_update;
+
+};