ClearSpeed / Application Story -How Jungo helps ClearSpeed in delivering robust products faster

ClearSpeed's processors have been designed for use in applications that are data-parallel in nature. Such data-parallel problems are found in many different application areas especially High Performance Compute (HPC), engineering and scientific applications.

The Advance™ accelerator board has two of ClearSpeed's CSX600 processors, DRAM and a PCI-X interface. This can be used to accelerate math-intensive applications by offloading the processing to standard math libraries running on the Advance board. To do this, data is transferred to the accelerator board where it is processed by the CSX600 processors and the results are then returned to the host processor.

As well as standard application libraries such as BLAS, LAPACK and FFT, ClearSpeed provides a full set of development tools based around an optimizing C compiler and the industry-standard gdb debugger.

In order to take full advantage of the 64-bit architectures that dominate the target application areas of HPC and scientific computation, ClearSpeed required driver solutions to access its Advance PCI-X based board. ClearSpeed requested that Jungo port its leading WinDriver toolkit to support the 64-bit architecture as well, allowing ClearSpeed to use the same driver code on both Linux and Windows and on both 32-bit and 64-bit systems.

Jungo offers its Professional Services Team to add custom features to the WinDriver toolkit and integrate the WinDriver kernel module into additional environments. With extensive knowledge in operating systems and driver development Jungo's team was executed to conduct the integration and completed the project within a month time of development.

"WinDriver provides a cross platform PCI driver stack for both Windows and Linux platforms" said Con Bradley from ClearSpeed Technology who chose WinDriver as the foundation for the CSX600 drivers, "being also able to develop the drivers in user space was a big advantage as it was much easier to debug the drivers. Also driver bugs were less likely to cause a kernel crash."

WinDriver allowed ClearSpeed to focus on its core expertise and let Jungo handle the OS internals. Jungo's mature and rigorously tested kernel module could be used with the same API on multiple OSs and CPU architecture.

WinDriver API was used to map the address spaces on the CSX600 card, handle interrupts and perform DMA.

Performance was a critical issue within the ClearSpeed design, as the Advance accelerator board is geared to accelerate data computations. ClearSpeed required very high DMA performance and were able to achieve this by using WinDriver Kernel Module although accessing it from a user level application. Usin WinDriver advance DMA API, ClearSpeed was able to interface the DMA locking functions and set up a raw I/O DMA mechanism between the Advance accelerator board and the user level buffers within the application. WinDriver's optimized kernel module was able to supply the performance required by the ClearSpeed environment.

Using a single ClearSpeed Advance board, the performance of a Dual Opteron, IBM A Pro e-Server was increased from just over 8 GFLOPS to a breathtaking 32.2 GFLOPS while actually decreasing the energy consumption. This why the Tokyo Institute of Technology selected ClearSpeed for its new Sun based supercomputer which is planned to be one of the top ten largest super computers in the world.