HDF5 – concurrency, compression & I/O performance [closed]

Posted on

Question :

HDF5 – concurrency, compression & I/O performance [closed]

I have the following questions about HDF5 performance and concurrency:

  1. Does HDF5 support concurrent write access?
  2. Concurrency considerations aside, how is HDF5 performance in terms of I/O performance (does compression rates affect the performance)?
  3. Since I use HDF5 with Python, how does its performance compare to Sqlite?

References:

Answer #1:

Updated to use pandas 0.13.1

1) No. http://pandas.pydata.org/pandas-docs/dev/io.html#notes-caveats. There are various ways to do this, e.g. have your different threads/processes write out the computation results, then have a single process combine.

2) depending the type of data you store, how you do it, and how you want to retrieve, HDF5 can offer vastly better performance. Storing in an HDFStore as a single array, float data, compressed (in other words, not storing it in a format that allows for querying), will be stored/read amazing fast. Even storing in the table format (which slows down the write performance), will offer quite good write performance. You can look at this for some detailed comparsions (which is what HDFStore uses under the hood). http://www.pytables.org/, here’s a nice picture:

(and since PyTables 2.3 the queries are now indexed), so perf actually is MUCH better than this
So to answer your question, if you want any kind of performance, HDF5 is the way to go.

Writing:

In [14]: %timeit test_sql_write(df)
1 loops, best of 3: 6.24 s per loop

In [15]: %timeit test_hdf_fixed_write(df)
1 loops, best of 3: 237 ms per loop

In [16]: %timeit test_hdf_table_write(df)
1 loops, best of 3: 901 ms per loop

In [17]: %timeit test_csv_write(df)
1 loops, best of 3: 3.44 s per loop

Reading

In [18]: %timeit test_sql_read()
1 loops, best of 3: 766 ms per loop

In [19]: %timeit test_hdf_fixed_read()
10 loops, best of 3: 19.1 ms per loop

In [20]: %timeit test_hdf_table_read()
10 loops, best of 3: 39 ms per loop

In [22]: %timeit test_csv_read()
1 loops, best of 3: 620 ms per loop

And here’s the code

import sqlite3
import os
from pandas.io import sql

In [3]: df = DataFrame(randn(1000000,2),columns=list('AB'))
<class 'pandas.core.frame.DataFrame'>
Int64Index: 1000000 entries, 0 to 999999
Data columns (total 2 columns):
A    1000000  non-null values
B    1000000  non-null values
dtypes: float64(2)

def test_sql_write(df):
    if os.path.exists('test.sql'):
        os.remove('test.sql')
    sql_db = sqlite3.connect('test.sql')
    sql.write_frame(df, name='test_table', con=sql_db)
    sql_db.close()

def test_sql_read():
    sql_db = sqlite3.connect('test.sql')
    sql.read_frame("select * from test_table", sql_db)
    sql_db.close()

def test_hdf_fixed_write(df):
    df.to_hdf('test_fixed.hdf','test',mode='w')

def test_csv_read():
    pd.read_csv('test.csv',index_col=0)

def test_csv_write(df):
    df.to_csv('test.csv',mode='w')    

def test_hdf_fixed_read():
    pd.read_hdf('test_fixed.hdf','test')

def test_hdf_table_write(df):
    df.to_hdf('test_table.hdf','test',format='table',mode='w')

def test_hdf_table_read():
    pd.read_hdf('test_table.hdf','test')

Of course YMMV.

Answered By: Jeff

Answer #2:

Look into pytables, they might have already done a lot of this legwork for you.

That said, I am not fully clear on how to compare hdf and sqlite. hdf is a general purpose hierarchical data file format + libraries and sqlite is a relational database.

hdf does support parallel I/O at the c level, but I am not sure how much of that h5py wraps or if it will play nice with NFS.

If you really want a highly concurrent relational database, why not just use a real SQL server?

Answered By: tacaswell

Leave a Reply

Your email address will not be published.