Topics

Charm++ is a machine independent parallel programming system. Programs written using this system will run unchanged on MIMD machines with or without a shared memory. It provides high-level mechanisms and strategies to facilitate the task of developing even highly complex parallel applications. Molecular dynamics (MD) simulation is a form of computer simulation in which atoms and molecules are allowed to interact for a period of time by approximations of known physics, giving a view of the motion of the particles. There are a lot of parallel methods applied to such application which needs great power of computing resources. NAMD is a state-of-art parallel molecular dynamics code designed for high-performance simulation of large bio-molecular systems based on charm++ parallel language. Currently, NAMD scales to hundreds of processors on high-end parallel platforms and tens of processors on commodity clusters using gigabit ethernet. (more information)

GPU is dedicated processor for rendering graphics. There are two dominant producers of High Performance GPUs chips: NVIDIA and ATM. Most GPU programs are written in shader language such as OpenGL (Linux, Windows) or HLSL (Windows). FPGAs have a long history in embedded processing and specialized computing. These areas include DSP, ASIC prototyping, medical imaging and other specialized compute intensive areas. An important differentiator between FPGAs and other accelerators is that they are programmable. The dominant FPGA chip vendors are Xilinx and Altera. ClearSpeed Technology produces a board that is designed to accelerate floating-point calculations. This board plugs into a PCI-X slot, has a clock cycle of 500 MHz, and contains 96 floating-point functional units that can each perform a double precision multiply-add in one cycle.

Draw shortly the differences between the two architecture, putting the emphasis on the advantages and limitation of CUDA. List performance tuning techniques for CUDA with supportive explanation of each. Performance techniques should constitute at least 60% of your report/presentation.

Also the performance analysis procedure itself is more difficult. This seminar topic will provide an insight into performance analysis of general purpose applications running on GPUs (GPGPU), performance analysis tools available and important optimization issues.

To answer the need of more and more computational power the underlying hardware gets more and more parallel. Nowadays the number of computational cores on each chip is growing fast and heterogenous systems are used soaring. This also affects the area of High Performance Computing, where parallelization on clusters or supercomputers was already in heavy use for many years.

Parallelism on the hardware level can be found in many layers. Almost all general purpose processing units are using several functional units in one core to be able to finish more operations per clock cycle. To make use of these capabilities the processor includes scheduling and reordering of instructions which is heavily influenced by the dependencies of these instructions.

Logical independent execution streams can be scheduled as separate threads and therefore optimize the usage of the available functional units and pipelines on multithreaded CPUs.

Within these functional units you can find another level of parallelism like VLIW or SWAR to apply computational operations on multiple words.

Current trends in CPU architectures are adding more and more cores on a single chip. They can be quite different in their capabilities or connection to some on-chip-network for communication with other cores and the memory.

The next level combines multiple chips in one machine, in the case of smaller machines many of them in a cluster and ?nally even clustering several clusters over the internet which must also be addressed by a programming model.


There are many different approaches on how to fetch, compute and manage the data within a processor. SIMD, MIMD, SPMD, Vector Processing or Stream Processing are some common processing models – each with its special advantages and disadvantages. Processing models expose the functionality of the processor and are specified by the vendor. They are optimized for performance and simplicity.