Programming for the Linux Cluster

The primary purpose of the SCV Linux Cluster is to run multiprocessor applications. The system consists of 1 login nodes, cootie.bu.edu, and 30 compute nodes. The login node is for code development and compilation, while the compute nodes are dedicated to serial and multiprocessor batch processing.

Before proceeding to compile your code, be sure to go to the Code Porting page first to see if there are any hardware or software constraints that might affect your code. If your code requires external packages (e.g., linear algebra subroutines ), go to the Packages page to see what is available. On the Cluster, all multiprocessor jobs must be submitted to the batch system to run on the back-end. This topic is covered in the Running Jobs page.

Compile  

The primary means for multiprocessing on the Cluster is MPI. Technically, OpenMP is also supported. However, because shared-memory support is limited to the 2 processors within each node, running a code parallelized with OpenMP alone is not practical.

There are several C and Fortran compilers from different vendors available on the Cluster. Overall, they perform comparably, both in single processor and multiprocessor (MPI) modes. Hence, the compiler choice should probably be based on your familiarity and preference, rather than on performance grounds. However, if you use a compiler's automatic parallelization feature, it is possible that one compiler may be more effective than another. This is highly code-dependent.

In addition, there may be more than one version of a vendor's compiler available on the system at any given time. To see which versions of compilers are currently selected by the system, use the command:

cootie:~ % module list
Currently Loaded Modulefiles:
  1) gcc/3.4.5(default)
  2) intel/9.0(default)
  3) mpichgm/1.2.6..14b_gcc(default:mpich-gcc)

For users who prefer to use the system's default compilers, no action is required. For users who prefer to use a specific version of a compiler, there are three approaches.

If your code does not require a specific compiler, then the first approach is to simply use the system default compilers (i.e., GNU compilers). In this case, you don't have to take any action. On the other hand, if you must use a particular compiler, then you may want to consider one of the two approaches below.

The third approach is to override the system default via the module command. The advantage of this method, especially if you make the selection via the module command in your starting shell script, is essentially that you picked your own default compiler and no absolute path needs to be provided.

Shown below is a table which includes all of the available compilers with details on two key parallel features: OpenMP and MPI.

Scratch Space Usage  

In regards to using scratch space, you should use only the local scratch whenever possible as it will be more efficient than accessing scratch of another node. Here are some additional details on /scratch:

• If you are on a node, a reference to /scratch automatically points to the local node's /scratch, which is equivalent to /net/`hostname`/scratch.
• From any of the compute nodes, you will NOT be able to access either Skate or Cootie's scratch.
• You CANNOT access cootie scratch from skate or vice versa.
• To access cootie or skate's own (local) scratch, type "cd /scratch" at system prompt.

Programming Tools

• Debuggers
• Gnu (serial) debugger (gdb, Document)
• Intel (serial and parallel) debugger (idb, Manual, Release Notes)
• Profilers
• Gnu profiler (gprof)