jbd / pwalk Goto Github PK

About:
------

parallelwalk.py: a parallel directory traversal library. The library implements
the algorithm described by LaFon, Misra and Bringhurst.

http://conferences.computer.org/sc/2012/papers/1000a015.pdf

The following tools are build on top of parallelwalk.py:

pdu: a parallel du implementation.

pmd5: a parallel directory comparision tool.

Latest Version:
---------------

The latest version of pdu can be downloaded from 

https://github.com/wtsi-ssg/pwalk


Getting Started
---------------

pdu will summarise the disk usage and number of files for a directory. Its 
output is similar to du -s. You can use the -h flag to print out disk usage is human readable 
units.

pdu should be invoked via mpirun (although it will run with a single process).

mpirun -n10 pdu -h /my/big/directory

The optimal amount of parallelism will depend on your filesystem implementation.
Lustre (as of version 2.5) has a per-client metadata limit and so you will see increased 
performance if you tell MPI to spread processes across a number of different clients. 

2 processes per host and 10 hosts is a good starting point.


Extending parallelwalk.py
-------------------------

parallelwalk provides a basic class that will walk the filesystem, but carry out no actions on
files that it encounters. It should be easy to extend the class for use in your own programs. 
See the pydoc documentation in the library for the details.