Updated parser script and readme (#71)

* Updated to NV internal version 5.2 (5d662c51)

19 Nov, 2020 
Changelist for v5.2
- Fixed bugs related to naming convention in graphing function (power data was being stored into a loadgen variable).
- Made clearer that a workload name is needed in the loadgen directory structure and will warn the user
-- However, if *summary.txt or *detail.txt is by itself, there is currently no warning in place that the data parse is skipped


18 Nov, 2020 
Changelist for v5.1
- Added energy calculation to data sets that have "watts" or "power" in the label name.
- *Note* unit labels are not handled and the user must be aware and responsible for the units of the data


15 Sep, 2020 
Changelist for v5 
- Bug fix to graph filtering by run ID
- Bug fix to stats parsing
- Graph function polishing
- Expanded stats & graph window controls (as internal variable controls)
- Added deskew to command-line parameters.  Allows for manual timing adjustments.
- Added verbose to command-line parameters.  Moved most script messages to this flag.  Only error messages will be display be default.
- Added stats to command-line parameters.  If data is a number format, outputs data statistics to stdout.  Optional list of case-sensitive strings sub-parameter to filter for specific data.
- Added csv to command-line parameters.  Used with stats parameter to save statistics to a CSV file.  Optional filename sub-parameter.
- Updated graph command-line parameter to take in an optional list of case-sensitive strings as a sub-parameter to filter for specific data.
- Added workload to command-line parameters.  Requires a list of case-sensitive list of strings to be used as either a filter or to specify an unsupported workload.  Default list of workloads: resnet, ssd-large, ssd-small, mobilenet, gnmt.


01 Jun, 2020
Changelist for v4 (let's skip v3)
- Moved to using Pandas DataFrames for better data handling and statistical calculations.
- Graphing can now handle any set of CSV data.  However, explicit 'Date' and 'Time' columns are required.
- The DASH webpage is now more interactive and responds to various user inputs.
  * Filtering by keywords (OR'ed) AND'ed run ID (OR'ed) affects all figures
  * Hiding/showing traces via the legend box triggers updates to the Lodagen stats table
  * Using selection tool (box or lasso) will generate statistics for the selected points
- pytz has been dropped in favor of a user-specified timedelta setting
- Loadgen CSV generation has been modified to output a proper CSV table.
  Previous CSVs generated by older versions of this script will no longer work with the graphing function.


21 Apr, 2020
Changelist for v2
- Changed date/time to unified, iso format (YYYY-mm-DD HH:MM:SS.fff)
- Added graphing feature (use -g/--graph)

* Updated to reflect v5

* Update parse_mlperf.py to v6

Changelist in v6:
- Removed ability to adjust statistical window for calculations
- Added colored window to highlight loadgen window found and used for statistics (plotly v4.12 or later required)

* Updated to reflect v6

* Update parse_mlperf.py

Minor bug fix:
- Fixed issue where loadgen timestamps are present but no associated data for drawing the loadgen window.  Therefore both loadgen and data need to be present

log_parsers/README.md

@@ -3,62 +3,96 @@
 
 # Dependencies
 
-Developed under Python 3.
+Developed under Python 3 for Windows.
+Other OS enviroments should work.
+
+Latest versions should work, but not actively tested.  The versions below were used for development.
 
 The graphing feature uses plotly.
 To install:
 ```
-  pip install plotly==4.6.0
+  pip install dash==1.18.1
+  pip install plotly==4.14.1
+```
+
+Data handling uses pandas & numpy
+To install:
+```
+  pip install pandas==1.0.5
+  pip install numpy==1.19.1
 ```
 
-Timezone adjustment features uses pytz
+Date parsing uses dateutils
 To install:
 ```
-  pip install pytz
+  pip install dateutil
 ```
 
+
 # Script In-Line Paramters
 
 Inside the parser script are some global variables/options.
 
-The following variables are for timing offsets between Host (PTDaemon, usually uses local time) and DUT (usually in UTC).
-```
-  g_power_tz               = None # pytz.timezone( 'US/Pacific' )
-                             # Refer to pytz for timezone list.  This sets the timezone for the Host system
-                             
-  g_loadgen_tz             = None # pytz.utc
-                             # Refer to pytz for timezone list.  This sets the timezone for the DUT system.  Typically does not need to be set.
-                             
-  g_power_add_td           = timedelta(seconds=3600)
-                             # This parameter will add the variable of seconds to the timedelta between loadgen and powerlog timestamps
-                             
-  g_power_sub_td           = timedelta(seconds=0)
-                             # This parameter will subtract the variable of seconds to the timedelta between loadgen and powerlog timestamps
-                             # This is b/c timedelta does not use negatives in the seconds place, only in the days place.
-```
-
 The following variables are for modifying the graphing and statistical windows.
 ```
-  g_power_window_before_td = timedelta(seconds=30)
-  g_power_window_after_td  = timedelta(seconds=30)
-                             # These parameters will collate additional data the variable of seconds BEFORE loadgen's BEGIN time
-                             # and the variable of seconds AFTER loadgen's END time
-                             
-  g_power_stats_begin_td   = timedelta(seconds=3)    # not implemented yet
-  g_power_stats_end_td     = timedelta(seconds=3)    # not implemented yet
+# g_power_window* : adjusts the time around POWER_BEGIN and POWER_END of the loadgen timestamps to show data in graph.
+                    typically used to hide or show setup or settling behavior for further analysis
+  g_power_window_before_add_td = timedelta(seconds=0)
+  g_power_window_before_sub_td = timedelta(seconds=0)
+  g_power_window_after_add_td  = timedelta(seconds=0)
+  g_power_window_after_sub_td  = timedelta(seconds=10)
 ```
 
 # Command-line Parameters
 
 ```
-  -lgi, --loadgen_in  : Directory of loadgen log files to parse
-  -pli, --power_in    : PTDaemon power log file to parse
-  -lgo, --loadgen_out : loadgen CSV output filename to write to
-  -plo, --power_out   : power log CSV output filename to write to
-  -g,   --graph       : graph the data, if possible.  Uses lgo and plo filenames as input
+  -h, --help            show this help message and exit
+  
+  -lgi LOADGEN_IN, --loadgen_in LOADGEN_IN
+                        Specify directory of loadgen log files to parase from
+  -spl SPECPOWER_IN, --specpower_in SPECPOWER_IN
+                        Specify PTDaemon power log file (in custom PTD format)
+  -pli POWERLOG_IN, --powerlog_in POWERLOG_IN
+                        Specify power or data input file (in CSV format)
+
+  -lgo LOADGEN_OUT, --loadgen_out LOADGEN_OUT
+                        Specify loadgen CSV output file (default:
+                        loadgen_out.csv)
+  -plo POWERLOG_OUT, --powerlog_out POWERLOG_OUT
+                        Specify power or data CSV output file (default:
+                        power_out.csv)
+
+  -g [GRAPH [GRAPH ...]], --graph [GRAPH [GRAPH ...]]
+                        Draw/output graphable data over time using the lgi/lgo
+                        and pli/plo as input. (Optional) Input a list of
+                        strings to filter data
+  -s [STATS [STATS ...]], --stats [STATS [STATS ...]]
+                        Outputs statistics between loadgen & power/data
+                        timestamps using lgi/lgo and pli/plo as inputs.
+                        (Optional) Input a list of strings to filter data
+  -csv [CSV], --csv [CSV]
+                        Outputs statistics to a CSV file (optional parameter,
+                        default: stats_out.csv) instead of stdout.
+  -w WORKLOAD [WORKLOAD ...], --workload WORKLOAD [WORKLOAD ...]
+                        Parse for workloads other than [mobilenet, gnmt,
+                        resenet50|resnet, ssd-large|ssdresnet34, or ssd-
+                        small|ssdmobilenet]
+
+  -v, --verbose
+
+  -deskew DESKEW, --deskew DESKEW
+                        Adjust timing skew between loadgen and power/data logs
+                        (in seconds)
 ```
 
+# Graph/Plots
+
+When using the graph (**-g**) option, the script will loop into server mode.  
+Use a browser to connect to http://localhost:8050 (if running on the same system) or to the IP of the system running the script to view the graph(s).
+
+To terminate the server (and script), press Ctrl-C (or equivalent).
+
+
 # Future plans
 
-- Incorporate Dash (from plotly) for better presentation of data and statistics
-- Maybe move global variables to command-line
+- Possible performance enhancements