Submitted 1-18-2016
Use multiple processes and threads on either the Windows or Linux operating systems and measure the performance and efficiency of each mechanism.
The assignment is composed of 3 separate small programs that perform the same computational and I/O intensive task the same number of times. They differ in that:
-
Program 1 is single threaded and performs the computation a number of times serially.
-
Program 2 is multi-threaded and performs the computation a number of times in parallel using threads.
-
Program 3 is multi-process and performs the computation a number of times in parallel using processes.
START TIMER
FOR EACH n in num\_times
Perform\_complex\_operation()
WRITE RESULT TO FILE
END FOR
END TIMER
PRINT TIME
START TIMER
FOR EACH n in num\_threads
START THREAD -\> Perform\_complex\_operation()
WRITE RESULT TO FILE
END FOR
FOR EACH n in num\_threads
WAIT FOR THREAD/JOIN
END FOR
END TIMER
PRINT TIME
START TIMER
FOR EACH n in num\_processes
START PROCESS -\> Perform\_complex\_operation()
WRITE RESULT TO FILE
END FOR
FOR EACH n in num\_processes
WAIT FOR processes
END FOR
END TIMER
PRINT TIME
The project was implemented in Python 2.7 and results generated on a Windows 10 Environment. Three python programs were written. Single.py, thread.py and process.py. For I/O activity each program wrote the results of the computational operation to an equivalently named “.txt” file in the working directory. Each program printed out the total time taken to run the requested number of operations in seconds.
The computational operation used is identical in each program to allow the results to be compared. It is adapted from the floating point prime decomposition code example provided from the following:
http://rosettacode.org/wiki/Prime_decomposition#Python:_Using_floating_point
The main changes to the example code were to hardcode the parameters and have the results written to a file.
The test system details are as follows:
CPU: Intel I5-2500K set at 4.1ghz.
Cores: 4 Cores and 4 Logical Threads (No Hyperthreading).
RAM: 8GB DDR-1300 dual channel.
HDD: Intel 320 SSD
OS: Windows 10 x64 Pro
Each program was run multiple times increasing the number of operations performed and the time recorded. The inbuilt timer function was verified by also doing several test runs utilizing the “Measure-Command” in Windows PowerShell to verify the time taken. The time reported from the programs and the Measure-Command were within milliseconds of each other with the “Measure-Command” being higher accounting for its slight additional overheard. Because of this it is assumed the programs self-reported computation times are accurate for the purpose of this test.
| Time in seconds to complete N operations. | |||||||
|---|---|---|---|---|---|---|---|
| 1 | 2 | 4 | 10 | 50 | 100 | 500 | |
| single | 0.3710000515 | 0.7170000076 | 1.4270000458 | 3.5880000591 | 17.8599998951 | 35.8869998455 | 179.1630001070 |
| threaded | 0.3619999886 | 0.7200000286 | 1.4429998398 | 3.6059999466 | 17.8910000324 | 35.7430000305 | 180.3659999370 |
| process | 0.4100000858 | 0.4119999409 | 0.4260001183 | 1.0420000553 | 5.0160000324 | 10.1750001907 | 49.6420001984 |
As can be seen from the above results there was no improvement in the time taken to complete the operations when using Python multi-threading in windows in fact the time taken is basically identical. As well I observed identical CPU usage (36%) between the single and multithreaded programs. The multi-process version however showed considerably different behaviour. When only performing one operation is it take a small but noticeable increase in time to complete. This is likely from the additional overhead of starting a processes. However there is no significant increase between running 1 – 4 threads. This is likely due the to the processor having 4 cores. Above 4 operations the time to complete the program does increase as each core is now running multiple processes however the program completes orders of magnitude faster than the single and multi-threaded versions. It appears only the muti-processes version of the program can take advantage of the multiple cores on the machine. Indeed observed CPU usage maxed out at 100% compared to 36% for the other versions. Ram usage was also significantly higher.
By simple observation we can see that the multi-process version is faster because it actually launches multiple instances of the python interpreter which can then be managed by the OS scheduler to run simultaneous across all available hardware threads/processors. Indeed when the multi-process version of the program is run an equivalent number of python interpreter process entries can be seen in the windows task manager. This allows the program to complete much quicker at the expense of maxing out CPU usage and much higher ram usage.
To find out why the multi-threaded version of the program acted identically to the single threaded version I was required to perform some research. According to the following:
https://wiki.python.org/moin/GlobalInterpreterLock
The python interpreter is not inherently threadsafe. So when using multiple threads are used it implements a Global Interpreter Lock. In the case of this program the lock prevents any benefits of multithreading to be realised. Even in situations where the GIL is not locking the program is it known to add significant overhead.
Extract the submitted zip file on the respective host machines and navigate to the “App” folder. Ensure python is installed on the host machine and has been added to the Environment PATH. The program is run with the following commands where “n” is the number of operations to be performed. Results will be written to corresponding txt files in the same directory.
python single.py n
python thread.py n
python process.py n
This has was tested in a Windows environment however it should work in linux and OSX as well.