Introduction

libpninexus is a C++ implementation of the Nexus file format. Althoug the NIAC provides a C++ interface to their C-API too, libpninexus is a pure C++ implementation providing all the ideoms a C++ programmer would expect from an API. In contrary to the NIAC implementation libpninexus supports only the HDF5 file format as a backend to store data (this is virtually its only restriction in comparison to the existing C++ interface). Aside from this libpninexus provides the following advantages over the existing API

- Support for reading and writing \c std::string objects of variable length
- keeping multiple object open for reading and writing concurrently (in the
  original NIAC API an object must be explicitly opened before data can be
  written or read from it)
- support for complex numbers (32, 64, and 128 Bit) and for 128 Bit floating
  point numbers

Outlook - concurrent applications

As demands on computing power increase in PNI experiments applications dealing with data will run concurrently. Their are two methods to let an application run in parallel

threads - if an application runs on a single host
MPI for distributed cluster systems

While the later approach is well known in scientific computing, multithreaded applications are still rare. In order to write data from such applications efficiently libpninexus will support both approaches.

Thread safety

With the new standard C++11 C++ provides multithreading support over different platforms out of the box. However, HDF5 actually does not support multithreaded reading or writing. Indeed library calls are serialized using a global lock in the library. However, this is only true as long as the library is compiled with thread-safety switched on. If this is not the case an application has to take care about locking. libpninexus will do this job for you in future. Unfortunately it is not enough to simply introduced locks to make the code thread safe. Along with this also the interface has to be changed in order to reflect the problems that come with concurrency.

Support for MPI IO

The Message Passing Interface is a well known interface for writing applications running on distributed clusters. HDF5 supports parallel IO from such clusters under some restrictions. libpninexus will support the MPI interface of HDF5 to make it possible to read and write data in parallel.