Porting and Parallelising Codes

Choice of Machines

The CSAR machines have very similar architectures, all being shared memory machines built around SGI's NUMAlink interconnect. However, due to differences between the MIPS processors in the Origins, and the Itanium 2 processors in the Altix, the choice of machine can depend a lot on the type of code that you have and how much work you are prepared to put in (with our help) in modifying it so that it runs efficiently. You may wish to refer to the CSAR helpdesk about the suitability of the different machines for your type of code. When you have decided which machine you wish to run your code on then there are two aspects to consider for code development which are described below.

Types of Parallelism

The CSAR machines are all ccNUMA shared memory machines with an underlying distributed-memory architecture. ccNUMA stands for cache-coherent non-uniform memory access, and this means that the processors should all have the same view of memory at any given moment, but memory access times are not the same across the machine for any processor, with some parts of memory considered "closer" than others.

On a shared memory ccNUMA machine such as the Origin and Altix computers, parallel programs can be written in one of two ways:-

• Distributed Memory Model: In the distributed memory model, programs are written so that independent processes communicate with each other by inserting into the code calls to message passing libraries such as MPI or SHMEM. Computer codes using the MPI library are generally highly portable between different parallel machines.
• Shared Memory Model: Using the shared memory model, programs are parallelised by inserting compiler directives into the source files which tell the compiler to parallelise loops or sections of code. The most common, and portable, set of these directives is defined in the OpenMP standard.

Porting Codes

Apart from the programming paradigm, there are other issues that you should be aware of when porting a code between machines. The major ones are detailed below.

• Default Data Sizes: Different machines can have different default sizes for intrinsic numerical types such as a Fortran Real or a C float. For example on a Cray T3E a C float is 8 bytes long whereas on an SGI Origin it is 4 bytes long. If you do not take this into account, your program might still run without any changes but it might give incorrect results.
• Compilers: Compiler behaviour and compiler options can change between different systems and in some cases compiler flags can completely change their meaning. For example on the MIPSpro compiler on an Origin machine -mp indicates that there are OpenMP parallelisation directives in the code which should be compiled, whereas to the Intel compiler on the Altix the -mp option is an instruction to maintain precision by reducing the optimisation level.
• Non-standard Fortran Extensions: Often Fortran code is written to use functions or subroutines which are not part of the Fortran standard, but which are implemented by the compiler vendor. These routines might be timing routines, utilities to extend the input/output capabilities or extra numerical functions. For routines that are not part of the Fortran standard it may be necessary to use different function or subroutine calls on different machines.
• External Libraries: External libraries of routines which are used by the code may have different properties on different machines or may not even exist when moving from one architecture to another. Routine names may change, input arguments to functions may be different or completely new routines in different libraries may need to be selected.

CSAR Support