From 93c90edb6f7939efb103d0ca092de83fb25aae55 Mon Sep 17 00:00:00 2001
From: Volker Hoffmann <volker@cheleb.net>
Date: Tue, 12 Apr 2016 17:17:15 +0200
Subject: feat: add support for gpu stats from gtx cards

---
 Common/nvidia.py | 75 ++++++++++++++++++++++++++++++++++++--------------------
 ticker.py        |  4 +--
 2 files changed, 51 insertions(+), 28 deletions(-)

diff --git a/Common/nvidia.py b/Common/nvidia.py
index 88463b3..3579c7b 100644
--- a/Common/nvidia.py
+++ b/Common/nvidia.py
@@ -7,17 +7,23 @@ import numpy as np
 import glob as glob
 
 
-def read_gpu_stats(node='vesta1'):
+def read_gpu_stats(node='vesta1', type='tesla'):
     """
     Reads stats from CSV file, which was dumped on the GPU nodes.
     
-    This query generates the CSV file. Run on GPU node.
-    nvidia-smi --query-gpu=index,uuid,name,temperature.gpu,utilization.gpu,utilization.memory,power.draw --format=csv,noheader > vesta1_`date '+%s'`.csv
+    The following queries generate the CSV files.
+
+    Run on GPU node w/ Tesla Cards:
+    $ nvidia-smi --query-gpu=index,uuid,name,temperature.gpu,utilization.gpu,utilization.memory,power.draw --format=csv,noheader > vesta1_`date '+%s'`.csv
+
+    Run on GPU nodes w/ GTX Cards:
+    $ nvidia-smi --query-gpu=index,uuid,name,temperature.gpu --format=csv,noheader > tasna1_`date '+%s'`.csv
 
     CSV files is named "$node_$unix_epoch_in_seconds.csv".
     For example, "vesta1_1454748497.csv".
 
     @param: node - GPU node, used to nconstruct filename [String]
+    @param: type - GPU Type (Tesla|GTX) [String]
     @return: df - GPU Stats [Pandas Dataframe]
     @return: epoch - Time (Seconds since 01-01-1970) [Integer]
     @return: success - Did we get data? [Boolean]
@@ -42,29 +48,46 @@ def read_gpu_stats(node='vesta1'):
         # Extract Epoch
         epoch = int(fname[:-4].split('_')[-1])
 
-        names_cols = [ 'gpu_id', 'uuid', 'gpu_name', 'gpu_temperature', \
-                       'gpu_utilization', 'memory_utilization', 'power_draw' ]
-        df = pd.read_csv(fname, sep=', ', header=None, names=names_cols, \
-                         engine='python')
-
-        # Drop Units
-        df.power_draw = \
-            df.power_draw.apply(lambda a: a.split(' ')[0])
-        df.gpu_utilization = \
-            df.gpu_utilization.apply(lambda a: a.split(' ')[0])
-        df.memory_utilization = \
-            df.memory_utilization.apply(lambda a: a.split(' ')[0])
-        
-        # Fix Types
-        df.power_draw = \
-            np.asarray(df.power_draw, dtype=np.float64)
-        df.gpu_utilization = \
-            np.asarray(df.gpu_utilization, dtype=np.float64)
-        df.memory_utilization = \
-            np.asarray(df.memory_utilization, dtype=np.float64)
-
-        # Add Node Column
-        df['node'] = pd.Series(["%s" % node]*len(df))
+        # Branch on GPU Type. See header documentation.
+        # Tesla cards report all stats. GTX cards are gimped.
+        if type == 'tesla':
+            names_cols = [ 'gpu_id', 'uuid', \
+                           'gpu_name', \
+                           'gpu_temperature', \
+                           'gpu_utilization', \
+                           'memory_utilization', \
+                           'power_draw' ]
+            df = pd.read_csv(fname, sep=', ', header=None, names=names_cols, \
+                             engine='python')
+
+            # Drop Units
+            df.power_draw = \
+                df.power_draw.apply(lambda a: a.split(' ')[0])
+            df.gpu_utilization = \
+                df.gpu_utilization.apply(lambda a: a.split(' ')[0])
+            df.memory_utilization = \
+                df.memory_utilization.apply(lambda a: a.split(' ')[0])
+
+            # Fix Types
+            df.power_draw = \
+                np.asarray(df.power_draw, dtype=np.float64)
+            df.gpu_utilization = \
+                np.asarray(df.gpu_utilization, dtype=np.float64)
+            df.memory_utilization = \
+                np.asarray(df.memory_utilization, dtype=np.float64)
+
+            # Add Node Column
+            df['node'] = pd.Series(["%s" % node]*len(df))
+
+        elif type == 'gtx':
+            names_cols = [ 'gpu_id', 'uuid', \
+                           'gpu_name', \
+                           'gpu_temperature' ]
+            df = pd.read_csv(fname, sep=', ', header=None, names=names_cols, \
+                             engine='python')
+
+            # Add Node Column
+            df['node'] = pd.Series(["%s" % node]*len(df))
 
         # Return
         return df, epoch, True
diff --git a/ticker.py b/ticker.py
index 6b7d799..a7dd3ea 100644
--- a/ticker.py
+++ b/ticker.py
@@ -75,9 +75,9 @@ line = "room_temperature,room=zbox_room value=%.2f %i" % \
     (temperature, netatmo_epoch)
 lines.append(line)
 
-# GPU Stats
+# GPU Stats, Tesla Cards
 for gpu_node in [ 'vesta1', 'vesta2' ]:
-    df, gpu_epoch, sucess = nvidia.read_gpu_stats(node=gpu_node)
+    df, gpu_epoch, sucess = nvidia.read_gpu_stats(node=gpu_node, type='tesla')
     for irow, [ index, row ] in enumerate(df.iterrows()):
         line_01 = "gpu_temperature,node=%s,uuid=%s value=%.2f %i" % \
             (row.node, row.uuid, row.gpu_temperature, gpu_epoch)
-- 
cgit v1.1