refactor, two profiles

README.md

@@ -5,8 +5,9 @@ Hacky solution to integrate AMDGPU power control and overclocking in `tuned` wit
 Takes a list of existing `tuned` profiles and creates new ones based on them.  These new profiles include AMDGPU power/clock parameters
 
 An attempt is made to discover the active GPU using the 'connected' state in the `DRM` subsystem, example:
-```
-$ grep -ls ^connected /sys/class/drm/*/status | grep -o card[0-9] | sort | uniq | sort -h | tail -1
+
+```bash
+~ $ grep -ls ^connected /sys/class/drm/*/status | grep -o card[0-9] | sort | uniq | sort -h | tail -1
 card1
 ```
 

playbook.yml

@@ -7,15 +7,22 @@
     - role: tuned_amdgpu
       # note: 'gpu_*' vars only apply with the 'custom' suffixed profiles created by this tooling
       # profiles based on the 'default' amdgpu power profile mode use default clocks
-      gpu_clock_min: "500" # default 500
-      gpu_clock_max: "2615" # default 2529
+      #
+      # the connected AMD GPU is automatically discovered - assumes one
+      # on swap to other AMD cards to avoid instability:
+      #   'rm -rfv /etc/tuned/*amdgpu*'
+      gpu_clock_min: "2200" # default 500, for best performance: near maximum. applies with 'overclock' tuned profile
+      gpu_clock_max: "2725" # default somewhere around 2529 to 2660
       gpumem_clock_static: "1075"
       gpu_power_multi:
-        default: 0.789473684210526 # 255W - slightly reduced, maintains clocks well
-        overclock: 0.869969040247678 # 281W - real default, board supports up to 323W (1.0)
+        default: 0.869969040247678 # 281W - real default
+#        overclock: 0.928792569659443 # 300W - slight boost
+        overclock: 1.0 # 323W - full board capability
       # optional, applies offset (+/-) to GPU voltage by provided mV
       # gpu_mv_offset: "-25"
+      gu_mv_offset: "+75" # add 50mV or 0.075V
       # '-50' undervolts GPU core voltage 50mV or 0.05V
+      # mostly untested, there be dragons/instability
       #
       # list of source tuned profiles available on Fedora (TODO: should dynamically discover)
       base_profiles:
@@ -26,30 +33,3 @@
         - network-throughput
         - powersave
         - virtual-host
-      #
-      # mapping of typical Navi generation power profiles from:
-      #   /sys/class/drm/card*/device/pp_power_profile_mode
-      # ref: https://www.kernel.org/doc/html/v4.20/gpu/amdgpu.html#pp-power-profile-mode
-      # 'pwr_cap_multi' is multiplied against board *limit* to determine profile wattage; 0.5 = 50%
-      # values below reflect my 6900XT
-#      amdgpu_profiles:
-#        default:
-#          pwrmode: 0
-#          pwr_cap_multi: 0.789473684210526 # 255W - slightly reduced, maintains clocks well
-#        3D:
-#          pwrmode: 1
-#          pwr_cap_multi: 0.869969040247678 # 281W - default
-#        powersave:
-#          pwrmode: 2
-#          pwr_cap_multi: 0.869969040247678
-#        VR:
-#          pwrmode: 4
-#          pwr_cap_multi: 0.869969040247678
-#        compute:
-#          pwrmode: 5
-#          pwr_cap_multi: 0.869969040247678
-#        custom:
-#          pwrmode: 6
-#          pwr_cap_multi: 1.0 # 323W - full capability
-      # both dictionaries are merged to create new 'tuned' profiles. eg:
-      #   'balanced-amdgpu-default', 'balanced-amdgpu-3D', 'balanced-amdgpu-video'

roles/tuned_amdgpu/defaults/main.yml

@@ -1,27 +1,11 @@
 ---
 # defaults file for tuned_amdgpu
 #
-# vars handling unit conversion RE: power capabilities/limits
-# the discovered board limit for power capability; in microWatts, then converted
-power_max: "{{ power_max_b64['content'] | b64decode }}"
-board_watts: "{{ power_max | int / 1000000 }}"
 
 # internals for profile power calculations
 # item in the context of the with_nested loops in the play
 profile_name: "{{ item.0 }}"
 
-# determine percentage for human-friendly comments
-power_default_pct: "{{ (gpu_power_multi.default * 100.0) | round(2) }}"
-power_oc_pct: "{{ (gpu_power_multi.overclock * 100.0) | round(2) }}"
-
-# in microWatts, actually written to sysfs
-power_default_mw: "{{ (power_max | float) * (gpu_power_multi.default | float) }}"
-power_oc_mw: "{{ (power_max | float) * (gpu_power_multi.overclock | float) }}"
-
-# wattages - more human-friendly comments
-power_default_watts: "{{ (power_default_mw | int) / 1000000 }}"
-power_oc_watts: "{{ (power_oc_mw | int) / 1000000 }}"
-
 amdgpu_profiles:
   - default
   - overclock

roles/tuned_amdgpu/tasks/main.yml

@@ -28,38 +28,6 @@
   when: (fed_ppdtuned_swap is not defined) or ('tuned' not in ansible_facts.packages)
   become: true
 
-- name: Determine GPU device in drm subsystem
-  ansible.builtin.shell:
-    cmd: grep -ls ^connected /sys/class/drm/*/status | grep -o card[0-9] | sort | uniq | sort -h | tail -1
-    executable: /bin/bash
-  changed_when: false
-  register: card
-
-- name: Find hwmon/max power capability file for {{ card.stdout }}
-  ansible.builtin.find:
-    paths: /sys/class/drm/{{ card.stdout }}/device/hwmon
-    file_type: file
-    recurse: true
-    use_regex: true
-    patterns:
-      - '^power1_cap_max$'
-  register: hwmon
-
-- name: Find hwmon/current power limit file for {{ card.stdout }}
-  ansible.builtin.find:
-    paths: /sys/class/drm/{{ card.stdout }}/device/hwmon
-    file_type: file
-    recurse: true
-    use_regex: true
-    patterns:
-      - '^power1_cap$'
-  register: powercap_set
-
-- name: Get max power capability for {{ card.stdout }}
-  ansible.builtin.slurp:
-    src: "{{ hwmon.files.0.path }}"
-  register: power_max_b64
-
 - name: Create custom profile directories
   ansible.builtin.file:
     state: directory

roles/tuned_amdgpu/templates/amdgpu-clock.sh.j2

@@ -18,6 +18,29 @@
 # dynamically determine the connected GPU using the DRM subsystem
 CARD=$(/usr/bin/grep -ls ^connected /sys/class/drm/*/status | /usr/bin/grep -o 'card[0-9]' | /usr/bin/sort | /usr/bin/uniq | /usr/bin/sort -h | /usr/bin/tail -1)
 
+function get_hwmon_dir() {
+    CARD_DIR="/sys/class/drm/${1}/device/"
+    for CANDIDATE in "${CARD_DIR}"/hwmon/hwmon*; do
+            if [[ -f "${CANDIDATE}"/power1_cap ]]; then
+                    # found a valid hwmon dir
+                    echo "${CANDIDATE}"
+            fi
+    done
+}
+
+# determine the hwmon directory
+HWMON_DIR=$(get_hwmon_dir "${CARD}")
+
+# read all of the power profiles, used to get the IDs for assignment later
+PROFILE_MODES=$(< /sys/class/drm/"${CARD}"/device/pp_power_profile_mode)
+
+# get power capability; later used determine limits
+read -r -d '' POWER_CAP < "$HWMON_DIR"/power1_cap_max
+
+# enable THP; profile enables the 'vm.compaction_proactiveness' sysctl
+# improves allocation latency
+echo 'always' | tee /sys/kernel/mm/transparent_hugepage/enabled
+
 {# begin the templated script for 'default' profiles to reset state #}
 {% if 'default' in profile_name %}
 # set control mode back to auto
@@ -27,29 +50,27 @@ echo 'auto' | tee /sys/class/drm/"${CARD}"/device/power_dpm_force_performance_le
 # reset any existing profile clock changes
 echo 'r' | tee /sys/class/drm/"${CARD}"/device/pp_od_clk_voltage
 
-# give '{{ profile_name }}' profile ~{{ power_default_pct }}% (rounded) of the max power capability
-#   {{ power_default_watts }} Watts of {{ board_watts }} total
-echo '{{ power_default_mw | int }}' | tee '{{ powercap_set.files.0.path }}'
+# adjust power limit using multiplier against board capability
+POWER_LIM_DEFAULT=$(/usr/bin/awk -v m="$POWER_CAP" -v n={{ gpu_power_multi.default }} 'BEGIN {printf "%.0f", (m*n)}')
+echo "$POWER_LIM_DEFAULT" | tee "${HWMON_DIR}/power1_cap"
+
+# extract the power-saving profile ID number
+PROF_POWER_SAVING_NUM=$(/usr/bin/awk '$0 ~ /POWER_SAVING.*:/ {print $1}' <<< "$PROFILE_MODES")
+
+# reset power/clock heuristics to power-saving
+echo "${PROF_POWER_SAVING_NUM}" | tee /sys/class/drm/"${CARD}"/device/pp_power_profile_mode
 {% else %}
-{# begin the templated script for non-default AMD GPU profiles, eg: 'VR' or '3D_FULL_SCREEN' #}
-# set manual control mode
-#   allows control via 'pp_dpm_mclk', 'pp_dpm_sclk', 'pp_dpm_pcie', 'pp_dpm_fclk', and 'pp_power_profile_mode' files
-#   only interested in 'pp_power_profile_mode' for power and 'pp_dpm_mclk' for memory clock (flickering).
-#   GPU clocks are dynamic based on (load) condition
-#echo 'manual' | tee /sys/class/drm/"${CARD}"/device/power_dpm_force_performance_level
-
-# give '{{ profile_name }}' profile ~{{ power_oc_pct }}% (rounded) of the max power capability
-#   {{ power_oc_watts }} Watts of {{ board_watts }} total
-echo '{{ power_oc_mw | int }}' | tee '{{ powercap_set.files.0.path }}'
-
-# set the minimum GPU clock
+{# begin the templated script for 'overclocked' AMD GPU profiles based on the existing tuned profiles #}
+# set the minimum GPU clock - for best performance, this should be near the maximum
+# RX6000 series power management *sucks*
 echo 's 0 {{ gpu_clock_min }}' | tee /sys/class/drm/"${CARD}"/device/pp_od_clk_voltage
 
 # set the maximum GPU clock
 echo 's 1 {{ gpu_clock_max }}' | tee /sys/class/drm/"${CARD}"/device/pp_od_clk_voltage
 
-# set the minimum / maximum GPU *memory* clock - force it high
-echo 'm 0 {{ gpumem_clock_static }}' | tee /sys/class/drm/"${CARD}"/device/pp_od_clk_voltage
+# set the GPU *memory* clock
+# normally this would appear disregarded, memory clocked at the minimum allowed by the overdrive (OD) range
+# it follows the core clock; if both 0/1 profiles for _it_ are high enough, the memory will follow
 echo 'm 1 {{ gpumem_clock_static }}' | tee /sys/class/drm/"${CARD}"/device/pp_od_clk_voltage
 {% if gpu_mv_offset is defined %}
 
@@ -60,12 +81,30 @@ echo 'vo {{ gpu_mv_offset }}' | tee /sys/class/drm/"${CARD}"/device/pp_od_clk_vo
 # commit the changes
 echo 'c' | tee /sys/class/drm/"${CARD}"/device/pp_od_clk_voltage
 
-# force GPU memory into highest clock (fix flickering)
-#   pp_dpm_*clk settings are unintuitive, giving profiles that may be used
-#   opt not to set the others (eg: sclk/fclk) - those should remain for benefits from the curve
-# echo '3' | tee /sys/class/drm/"${CARD}"/device/pp_dpm_mclk
+# force GPU core and memory into highest clocks (fix flickering and poor power management)
+# set manual control mode
+#   allows control via 'pp_dpm_mclk', 'pp_dpm_sclk', 'pp_dpm_pcie', 'pp_dpm_fclk', and 'pp_power_profile_mode' files
+echo 'manual' | tee /sys/class/drm/"${CARD}"/device/power_dpm_force_performance_level
 
-# note 4/8/2023: instead of 'manual'... deal with broken power management, force clocks to high
+# adjust power limit using multiplier against board capability
+POWER_LIM_OC=$(/usr/bin/awk -v m="$POWER_CAP" -v n={{ gpu_power_multi.overclock }} 'BEGIN {printf "%.0f", (m*n)}')
+echo "$POWER_LIM_OC" | tee "${HWMON_DIR}/power1_cap"
+
+#   pp_dpm_*clk settings are unintuitive, giving profiles that may be used
+echo '1' | tee /sys/class/drm/"${CARD}"/device/pp_dpm_sclk
+echo '3' | tee /sys/class/drm/"${CARD}"/device/pp_dpm_mclk
+echo '2' | tee /sys/class/drm/"${CARD}"/device/pp_dpm_fclk
+echo '2' | tee /sys/class/drm/"${CARD}"/device/pp_dpm_socclk
+
+# extract the VR power profile ID number
+PROF_VR_NUM=$(/usr/bin/awk '$0 ~ /VR.*:/ {print $1}' <<< "$PROFILE_MODES")
+
+# force 'overclocked' profile to 'VR' power/clock heuristics
+# latency/frame timing seemed favorable with relatively-close minimum clocks
+echo "${PROF_VR_NUM}" | tee /sys/class/drm/"${CARD}"/device/pp_power_profile_mode
+
+# note 4/8/2023: instead of 'manual'... try dealing with broken power management, force clocks to high
 # ref: https://gitlab.freedesktop.org/drm/amd/-/issues/1500
-echo 'high' | tee /sys/class/drm/"${CARD}"/device/power_dpm_force_performance_level
+# followup: doesn't work that well in practice, still flaky on clocks/frame times
+#echo 'high' | tee /sys/class/drm/"${CARD}"/device/power_dpm_force_performance_level
 {% endif %}

roles/tuned_amdgpu/templates/tuned.conf.j2

@@ -11,8 +11,12 @@ net.core.default_qdisc=fq
 net.ipv4.tcp_congestion_control=bbr2
 net.core.rmem_max=33554432
 net.core.wmem_max=33554432
-dev.raid.speed_limit_min=600000
-dev.raid.speed_limit_max=9000000
+dev.raid.speed_limit_min=1000000
+dev.raid.speed_limit_max=6000000
+# improve THP allocation latency, compact in background
+vm.compaction_proactiveness=30
+# make page lock theft slightly more fair
+vm.page_lock_unfairness=1
 # allow some games to run (eg: DayZ)
 vm.max_map_count=1048576