No abstract

This article presents the design of a generic HyperTransport (HT) core. HyperTransport is a packetbased interconnect technology for low-latency, high-bandwidth point-to-point connections. It is specially optimized to achieve a very low latency. The core has been verified in system using an FPGA. This exhaustive verification and the generic design allow the mapping to both ASICs and FPGAs. The implementation described in this work supports a 16-bit link width, as used by Opteron processors. On a Xilinx Virtex-4 FX60, the core supports a link frequency of 400 MHz DDR and offers a maximum bidirectional bandwidth of 3.2GB/s. The in-system verification has been performed using a custom FPGA board that has been plugged into a HyperTransport extension connector (HTX) of a standard Opteron-based motherboard. HTX slots in Opteron-based motherboards allow very high-bandwidth, low-latency communication, since the HTX device is directly connected to one of the HyperTransport links of the processor. Performance analysis shows a unidirectional payload bandwidth of 1.4GB/s and a read latency of 180 ns. The HT core in combination with the HTX board is an ideal base for prototyping systems and implementing FPGA coprocessors. The HT core is available as open source.

No abstract

Although large scale high performance computing today typically relies on message passing, shared memory can offer significant advantages, as the overhead associated with MPI is completely avoided. In this way, we have developed an FPGA-based Shared Memory Engine that allows to forward memory transactions, like loads and stores, to remote memory locations in large clusters, thus providing a single memory address space. As coherency protocols do not scale with system size we completely avoid a global coherency across the cluster. However, we maintain local coherency domains, thus keeping the cores within one node coherent. In this paper, we show the suitability of our approach by analyzing the performance of barriers, a very common synchronization primitive in parallel programs. Experiments in a real cluster prototype show that our approach allows synchronization among 1024 cores spread over 64 nodes in less than 15us, several times faster than other highly optimized barriers. We show the feasibility of this approach by executing a shared-memory implementation of FFT. Finally, note that this barrier can also be leveraged by MPI applications running on our shared memory architecture for clusters. This ensures the usefulness of this work for applications already written.

No abstract