The GFI computing system

The Geophysical Institute has acquired a new computing system named cyclone.hpc.uib.no which replaces the system skd-cyclone.klientdrift.uib.no from Summer 2018.

The computing performance has been enhanced substantially: The new system is a Dell PowerEdge R740 Server configuration with the following characteristics:

Intel Xeon Gold 6140M, 18 cores, 72 threads, 2.3GHz, 25MB L3 cache

1.5 TB DDR4-2666 memory

2 NVIDIA Tesla GPU, 12 GB

Here are the key facts about using the new system

For transition, you will need to recompile your code (if it is compiled code) for the operating system CentOS linux. This is the same operating system as on the UiB high-performace compute system Hexagon.

There is no queue system on the new cyclone. Users can submit to the hexagon queue from cyclone (acts as login node).

There is a 30-50% resource limitation per user. This will keep individual users from accidentally brining down the system.

Different software configurations cand be activated using the command module

Access to data storage will be maintained using existing paths, such as /Data/gfi.

The same and additional software packages are available (your suggestions?)

The system is maintained by the experts from UiB's HPC group. This should be an advantage for GFI.

The new cyclone is equipped with very advanced computing power. There are two Tesla P100 Graphics Processing Units (GPUs) build into the server. This will boost parallel computation power for software that can make use of this feature. Native programming languages, standard libraries, as well as Matlab and python can make use of GPU acceleration for compute operations. GFI is planning on offering courses to introduce to GPU programming.

Here are the facts about the GPU accelerators:

each Tesla P100 has 3584 cores

memory of 12 GB per GPU

both GPUs can also be used in combination

programming using CUDA, DirectCompute, OpenCL, OpenACC

some libraries are already optimized for GPU usage

first optimisation step are OpenACC compiler statements

bottleneck in GPU programming is to get data on/off the GPU