NeIC 2015 accelerators workshop
Organizer: Ole Widar Saastad (USIT)
Reporting: Ole Widar Saastad (USIT)
Links
Accelerator status; present and in the future
Accelerators like NVIDIA CUDA and Intel MIC processors are increasingly relevant for production systems. Large scale production systems employ both NVDIA cuda and Intel MIC type processors (FPGA adaptation is too small to be considered in this workshop).
We know that major installations like Oak Ridge laboratory and Lawrence Livermore are pursuing the accelerator performance. While the theoretical performance is outstanding it has proven hard to utilise this performance for general production codes and even more so for legacy codes. Many of us are faced with a large fraction of such legacy codes and the need for tools like general purpose compilers are still unsatisfied. While openACC and openMP 4.0 are steps forward they still require change of the source code.
The workshop will focus on assessing the current situation around the sites, it would be beneficial if a representative from each relevant site could give a few minutes overview addressing the point below :
- Installed hardware, planned installment and planned procurement of accelerators.
- Current usage, applications using accelerators, tools like compilers etc.
- Work in progress to utilize accelerators using installed hardware and planned new hardware.
- Experience with accelerators.
The discussion should focus on how to address the situation with accelerators. Should future systems be fully accelerated or should accelerators be a small subset of exotic hardware attached ? The performance offered by the accelerators are tantalizing,and if harvested it would give significantly higher scientific production for comparatively less investment (at least in hardware).
The institutions like the US ones above are putting a lot of focus on tools that let the users have easy access to accelerator performance using standards like OpenACC, OpenMP 4.0 etc. Is this an option for Nordic sites ? If one decide to go all out using accelerators one must be prepared to do a major investment in tools, training and user support to exploit the accelerator hardware, including major rewrite of many applications.
On the other hard – are accelerators just a hype ? Will the future look like convergence between current x86-64 and MIC architectures ? Or maybe just more powerful MICs ? Or CUDA type massively threaded accelerators ?
Agenda
- Introduction by Ole W. Saastad, experiences at UiO
- Round table contribution from other sites reporting their experiences, please bring short presentations if you have something to share.
- Discussion about programming approaches and what scientists need - CUDA, OpenCL, OpenACC, OpenMP, MPI
- Future direction of accelerators, NVIDIA vs. standalone Intel MIC systems
- Next generation HPC systems - Accelerator based or not ? Partly accelerated ? Tools to offer scientists ? Training ?
Background reading