在python中存储大量数据

Question

Maybe I start will a small introduction for my problem. I'm writing a python program which will be used for post-processing of different physical simulations. Every simulation can create up to 100 GB of output. I deal with different informations (like positions, fields and densities,...) for different time steps. I would like to have the access to all this data at once which isn't possible because I don't have enough memory on my system. Normally I use read file and then do some operations and clear the memory. Then I read other data and do some operations and clear the memory.

Now my problem. If I do it this way, then I spend a lot of time to read data more than once. This take a lot of time. I would like to read it only once and store it for an easy access. Do you know a method to store a lot of data which is really fast or which doesn't need a lot of space.

I just created a method which is around ten times faster then a normal open-read. But I use cat (linux command) for that. It's a really dirty method and I would like to kick it out of my script.

Is it possible to use databases to store this data and to get the data faster than normal reading? (sorry for this question, but I'm not a computer scientist and I don't have a lot of knowledge behind databases).

EDIT:

My cat-code look something like this - only a example:

out = string.split(os.popen("cat "+base+"phs/phs01_00023_"+time).read())
# and if I want to have this data as arrays then I normally use and reshape (if I
# need it)
out = array(out)
out = reshape(out)

Normally I would use a numpy Method numpy.loadtxt which need the same time like normal reading.:

f = open('filename')
f.read()
...

I think that loadtxt just use the normal methods with some additional code lines.

I know there are some better ways to read out data. But everything what I found was really slow. I will now try mmap and hopefully I will have a better performance.

Answer 1

I would try using HDF5 . There are two commonly used Python interfaces, h5py and PyTables . While the latter seems to be more widespread, I prefer the former.

Answer 2

If you're on a 64-bit operating system, you can use the mmap module to map that entire file into memory space. Then, reading random bits of the data can be done a lot more quickly since the OS is then responsible for managing your access patterns. Note that you don't actually need 100 GB of RAM for this to work, since the OS will manage it all in virtual memory.

I've done this with a 30 GB file (the Wikipedia XML article dump) on 64-bit FreeBSD 8 with very good results.

Answer 3

If you're working with large datasets, Python may not be your best bet. If you want to use a database like MySQL or Postgres, you should give SQLAlchemy a try. It makes it quite easy to work with potentially large datasets using small Python objects. For example, if you use a definition like this:

from datetime import datetime
from sqlalchemy import Column, DateTime, Enum, ForeignKey, Integer, \
    MetaData, PickleType, String, Text, Table, LargeBinary
from sqlalchemy.ext.declarative import declarative_base
from sqlalchemy.orm import column_property, deferred, object_session, \
    relation, backref

SqlaBaseClass = declarative_base()

class MyDataObject(SqlaBaseClass):
  __tablename__ = 'datarows'
  eltid         = Column(Integer, primary_key=True)
  name          = Column(String(50, convert_unicode=True), nullable=False, unique=True, index=True)
  created       = Column(DateTime)
  updated       = Column(DateTime, default=datetime.today)

  mylargecontent = deferred(Column(LargeBinary))

  def __init__(self, name):
      self.name    = name
      self.created = datetime.today()

  def __repr__(self):
      return "<MyDataObject name='%s'>" %(self.name,)

Then you can easily access all rows using small data objects:

# set up database connection; open dbsession; ... 

for elt in dbsession.query(MyDataObject).all():
    print elt.eltid # does not access mylargecontent

    if (something(elt)):
        process(elt.mylargecontent) # now large binary is pulled from db server
                                    # on demand

I guess the point is: you can add as many fields to your data as you want, adding indexes as needed to speed up your search. And, most importantly, when you work with a MyDataObject , you can make potentially large fields deferred so that they are loaded only when you need them.