BALTRAD parallel processing

The default VM setup is to use a single CPU core. In order to demonstrate the power of parallel processing, you must first determine whether your physical hardware has more than a single core.

On Linux this is done in the terminal with the ‘nproc’ command.

On Mac this is done in the terminal with the ‘sysctl -n hw.ncpu’ command.

On Windows this is done graphically using the Task Manager’s Performance tab.

We want tune our VM to harness the power of several CPUs. Follow the following steps:

Shut down the IPython notebook Server (Ctrl-C, answer yes)
Shutdown the VM (click the X button in the VM window, choose power down the machine)
Select the VM in the VirtualBox Manager Window, from the menu choose Machine->Setting
Choose the System Tab, then Processor, use the slider to set the number of Processor to 2, 4, or 8 depending on your system resources.
Click Ok, and then start the machine
Login, use the start_notebook.sh script to start the IPython server, start the notebook and you should have multiple processors!

RELOAD THIS PAGE!

Verify from Python the number of CPU cores at our disposal

import multiprocessing
print("We have %i cores to play with!" % multiprocessing.cpu_count())

We have 4 cores to play with!

Yay! Now we’re going to set up some rudimentary functionality that will allow us to distribute a processing load among our cores.

Define a generator

import os
import _raveio, odc_polarQC

# Specify the processing chain
odc_polarQC.algorithm_ids = ["ropo", "beamb", "radvol-att", "radvol-broad", "rave-overshooting", "qi-total"]

# Run processing chain on a single file. Return an output file string.
def generate(file_string):
    rio = _raveio.open(file_string)

    pvol = rio.object
    pvol = odc_polarQC.QC(pvol)
    rio.object = pvol
    
    # Derive an output file name
    path, fstr = os.path.split(file_string)
    ofstr = os.path.join(path, 'qc_'+fstr)
    
    rio.save(ofstr)
    return ofstr

Feed the generator, sequentially

import glob, time

ifstrs = glob.glob("data/se*.h5")
before = time.time()
for fstr in ifstrs:
    print(fstr, generate(fstr))
after = time.time()

print("Processing time: %3.2f seconds" % (after-before))

data/sekkr.h5 data/qc_sekkr.h5

data/sevar.h5 data/qc_sevar.h5

data/sehud.h5 data/qc_sehud.h5

data/sease.h5 data/qc_sease.h5

data/selek.h5 data/qc_selek.h5

data/seang.h5 data/qc_seang.h5

data/seosu.h5 data/qc_seosu.h5

data/sevil.h5 data/qc_sevil.h5

data/selul.h5 data/qc_selul.h5

data/sekir.h5 data/qc_sekir.h5

data/searl.h5 data/qc_searl.h5
Processing time: 5.71 seconds

Mental note: repeat once!

Multiprocess the generator

# Both input and output are a list of file strings
def multi_generate(fstrs, procs=None):
    pool = multiprocessing.Pool(procs)  # Pool of processors. Defaults to all available logical cores

    results = []
    # chunksize=1 means feed a process a new job as soon as the process is idle.
    # In our case, this restricts the queue to one "dispatcher" which is faster.
    r = pool.map_async(generate, fstrs, chunksize=1, callback=results.append)
    r.wait()

    return results[0]

Feed the monster, asynchronously!

before = time.time()
ofstrs = multi_generate(ifstrs)
after = time.time()

print("Processing time: %3.2f seconds" % (after-before))

Processing time: 2.62 seconds