Lab 3

Lab 3, Python Profiling

We decided to not continue using MySQL chose to download python. We thought it would be a good idea to test it on both servers, australia: 64 bit x86 and red: 64 bit Arm.
We used python 3.4 using this link  and wget to get it onto the server. After we un-tarred it (tar -xvf [filename] ) we ran the ./configure which gave us a makefile and then we made a few changes.
We changed line 72 in the makefile which was
BASECFLAGS=         -Wno-unused-result
to
BASECFLAGS=         -Wno-unused-result -pg

This allows python to output a gmon.out file when python executes, this contains all the profiling information.
We now needed to run something with Python so we would have a valid gmon.out file.
we found this resource: tests which showed us how to run python tests using the command
./python -m test

On Australia:
This went through 389 tests when it was completed we had a gmon.out file
To make the profile easier to read our professor suggested to use gprof2dot to generate a png image file of the profile, the command is:
gprof ./python | gprof2dot | dot -Tpng > Profile.png  (full image)

  
On Red: 
This went through most of the tests but stalled at the last one with a message:
_mcleanup: gmon.out: No such file or directory 
(I later realized that message popped up a lot during the tests which suggests that it is not relevant to the stalling)
After waiting a long time I just used control-C to get out of the test and I still had a gmon.out with 388 tests. Again, used gprof2dot to get the image file of the profile (full image)

Although both pictures appear very different I was  looking at the function calls and they are the same, it is simply the layout of the picture that changes drastically.

There is one function that stands out on both systems is PyEval_EvalFrameEx, it takes up 12% on x86 and 14% on Arm. I looked up the source code here and the function itself is around 2500 lines , it is safe to say I'm not sure what it is doing. It does have an interesting comment section line 827 about optimizations that this function has, this is essentially trying to avoid leading the CPU on a mispredicted branch,  this relates to what Chris Tyler was talking about in our class(on 1/29/2015) when he mentioned CPUs guessing the correct path to follow.


In the end I think red's profile is hard to directly compare to australia's because of the difference of layout of the 2 files but the EvalFrameEx takes up the most usage on both systems. It seems like they are aware of this with the extensive optimization of it.

Thanks for reading