So far this is a pretty simple reader that just takes the binary output from ROSS and converts it to a CSV.
Just type 'make' when you're in the reader directory. If you're running on Linux on an x86 system, no flags needed. If you're running this on a BG/Q, you need to add the -DBGQ flag to the Makefile. If you're running this on a Mac, add -DMACOS, because Macs do not have the argp header needed for the runtime argument library used.
NOTE 0: You can't read data created on a system with a different endianness from the system you are running the reader on (i.e., don't read BG/Q data while running the reader on an x86 system).
NOTE 1: I haven't actually kept testing on Mac as I no longer do development on the Mac.
You need the ross-stats-README.txt file that is output
automatically by ROSS as it has some information that the reader needs.
It is expected to be in the same directory as your binary files.
Also if you're reading model level sampling data, you need an additional lp-mapping.txt file in the directory.
This is a csv of the format PE,KP,LP,LP_type
You can type ./ross-reader --help to see an explanation for the available options, which are also explained here.
-f or --filename is for inputting the filename of the binary file to be read.
-t or --filetype is for the type of file
File types:
- 0 GVT-based instrumentation
- 1 Real time sampling
- 2 event tracing
- 3 model level sampling data
The above arguments are required, the following are optional:
-c or --combine-events is for use with the event trace data. It will bin the event trace by
source, destination LP pairs over an interval of virtual time. It also gives the count of events
in that bin. This is useful if your event trace is really large and you'd like to have a smaller file
(at the cost of coarser grained data of course).
-b or --bin-size is to change the virtual time bin size. Default is 10000.
NOTE 2: Features related to these optional arguments are a work in progress. They may not work correctly.
Everything is output as a text file delimited by commas. The event tracing always outputs one file. If the ROSS instrumentation was selected to run with output only at a PE level, then only one file is output for the GVT-based data. Two are output for the real time, because the difference in virtual time and GVT metric is always reported at the KP level. So in this case, there will be a file for the PE metrics and one for the KP metrics. If the ROSS instrumentation was set to do the KP/LP level metrics, then both the GVT-based and real time output has one file each for PE, KP, and LP level metrics.
The text files are saved to the same directory as the binary files.
The reader needs the README file that is output by the instrumentation, because it contains data about the simulation the reader needs to correctly read the file. The reader expects that this file has the same naming convention as the data file (which it will, unless you change the file name manually).