Sparse matrix backends

«  Adding sample and observation metadata to biom files   ::   Contents   ::   Summarizing BIOM tables  »

Sparse matrix backends

As the BIOM project evolves, so do the underlying data structures, leading to potential runtime trade offs between implementations. Currently, there are three distinct sparse matrix backends to BIOM: the CSMat (default as of BIOM v1.1), the SparseMat and the SparseDict. Specific differences are discussed below.

How to check what sparse backend is in use

To check what sparse backend is in use, simply execute The last line shows the SparseObj type. For instance:

System information
      Platform: darwin
Python/GCC version: 2.7.2 (default, Mar 23 2012, 13:31:52)  [GCC 4.2.1 (Based on Apple Inc. build 5658) (LLVM build 2336.9.00)]
Python executable: /Users/mcdonald/bin/python

Dependency versions
              NumPy version:    1.6.1
biom-format library version:    1.1.0-dev
 biom-format script version:    1.1.0-dev

biom-format package information
SparseObj type: biom.csmat.CSMat

The last line indicates that the CSMat object is in use.

Changing the sparse backend

There are two methods to change the backend that is used. The first method is by copying the biom_config file located under support_files/ and placing it in your home directory as ~/.biom_config. Then, edit ``~/.biom_config and replace the current backend type with the desired type.

The second method is to set the environment variable $BIOM_CONFIG_FP to a file path of your choice, and place the following into that file:

python_code_sparse_backend      <BACKEND TYPE>

Where <BACKEND TYPE> is replaced by the specific backend implementation to use.

Sparse matrix backend descriptions

Different sparse matrix backends have different performance characteristics. As BIOM changes over time, additional methods may be added that address specific runtime concerns.


The default sparse backend is the CSMat. CSMat implements coordinate list, compressed sparse row and compressed sparse column formats and facilitates interaction with these representations. This backend will have the lowest memory footprint. In general, this backend should be the fastest. However, it has been observed that under certain circumstances, this backend may not perform the best.


The SparseMat is built using a combination of Python objects, a Cython wrapper and C++. It implements the dictionary of keys sparse matrix representation. This method performs pretty well, but has an increasing memory footprint as the number of nonzero values increases. Under some situations, specifically those that require a large number of Table creations, this backend has about a 15-20% reduced runtime over CSMat.


The SparseDict is the naive pure Python implementation. This was first implemented as a test backend, and it is not advised to use this object.

«  Adding sample and observation metadata to biom files   ::   Contents   ::   Summarizing BIOM tables  »