CSC requires all users to evaluate and report the scalability of their software in order to get permission for production runs on Louhi (louhi.csc.fi). The largest number of cores that a user can employ is decided basing on these results. This enables a fair and efficient usage of Louhi's resources. It is also very useful for the users to take some time to analyze the scalability and performance of their code, to utilize their CPU quotas optimally.

Scalability criteria

Louhi is intended for well-scaling parallel jobs. The minimum size of parallel jobs is 32 compute cores and the maximum size (without special arrangements) is 2048 compute cores. The current scalability criteria for Louhi are:

• When the number of cores is doubled  to 512, or less, the wall time must decrease by a factor of 1.5
• When the number of cores is doubled  from 512 to 1024, the wall time must decrease by a factor of 1.4
• When the number of cores is doubled  from 1024 to 2048, the wall time must decrease by a factor of 1.3

Scalability tests must be run through the batch job system. The data used for the test runs should reflect real production runs. The tests should start from a number of cores that is reasonable for the test data,  and they continue until the scalability limit has been reached (max 2048). The jobs have to be run with the same input data for each compute core count (parameters affecting the scalability can be and are encouraged to be changed).

For more details and for examples of batch job scripts, see:

• Louhi User's Guide

The scalability of the following programs have been measured, and can be referred to when asking for a permission for production runs on Louhi:

• CP2K 256 cores
• CPMD 256 cores
  
• DALTON 128 cores
• Elmfire 1024 cores
  
• Gadget-2 1024 cores
• GHA 4096 cores
  
• GPAW 1024 cores
  
• Gromacs 4.0  512 cores
• NAMD 512 cores
• Parcas 4096 cores
  
• Pencil 4096 cores
• SPATE-HPC (SUSWOOD) 2048 cores
  
• VASP 128 cores

For these program users do not need to perform any scalability tests, unless they wish to demonstrate better scalability for their specific dataset; but just making an announcement to helpdesk@csc.fi is sufficient.

Optional tests

We also encourage users to perform hardware performance counter measurements. There are detailed step-by-step instructions for performing these measurements on the following page:

http://www.csc.fi/tutkimus/Laskentapalvelut/laskenta/palvelimet/hw_perf_counters

Reporting

The scalability report should contain a short description of the program and the job as well as wall times with each compute core count. Especially, the factors that affect the scalability of the program should be elicited. The optional HW counter measurements should also be included if they have been performed. The report, together with a list of users whose quotas are affected by the tests, is then sent to helpdesk@csc.fi. The test account will be closed in a month if the results have not been reported. In case more time is needed, or any other questions have arisen, please contact  the same email address.