There will be a major system downtime starting 8:30am on Friday 19th August extending over the weekend. This will complete the NUMALink interconnect. After this it will be possible to run parallel jobs of up to 1536 processors.

The login node and 128 cpus will be available again later on Friday. The rest of the system will be back in production by Monday 22nd August and hopefully earlier.

The APAC board has decided that for the purposes of indicating the in-kind value of grants to bodies such as the ARC, a figure of $0.50 per processor hour on compute systems should be used.

For security reasons we ask that passwords are not communicated in emails or recorded where they can be easily seen.

• To ensure that OpenMP jobs are scheduled on the same shared memory segment you should always use -lncpus=N:N in your qsub request.
• OpenMP codes should be timed using differing numbers of threads to investigate the optimal number of threads for efficient use. There are several reasons for this. As the processors on the AC are faster than those of the SC the cost of starting up the threads can outweigh the cost of the computations in the loop. Make sure that you do as much computation as possible in each loop to balance the start-up cost. The shared memory programming style assumes uniform memory access across the entire memory address space. This is not true for NUMA architectures such as the Altix. To get good performance from OpenMP the code must ensure that data accessed by each process is local to the processor on which that process is currently executing. As the Altix uses a first-touch policy for data allocation you should ensure that data is initialised using the same distribution as it is going to have in later computations.

Most users expect their programs to crash (generate a floating point exception) in these circumstances. Forcing this behaviour on AC depends on which language you use:

• Fortran: Use the ifort compiler option -fpe0. Please read the ifort man page to understand the various -fpe options. On the AC system we have changed the default compiler behaviour from -fpe3 to -fpe0, so be aware of this when porting code.

• C/C++: Use the C99 floating point rounding and exception handling routines. See man fenv for details. As an example:

                   #define _GNU_SOURCE   // needed for icc but not icpc
                   #include < fenv.h >
  
                   ....
  
                   feenableexcept(FE_DIVBYZERO | FE_OVERFLOW);
               

If you would like any of these courses given at your partner site please contact us and it can be arranged. As these are hands-on courses we ask that a representative from the site arranges for access to a computer teaching lab with web access and a method of obtaining secure shell logins to the AC.