The cluster computing has become very popular in the world, which uses commodity components and achieves good cost/performance ratio for parallel computing. However, it is very hard for the conventional cluster computers to extract its potential performance due to its large overhead hided in the network hardware. Considering the cluster’s characteristics, the optimization of communication is available.
Maestro project proposed Network Burst and Pipelined transfer as the optimization technique oriented to cluster’s inter-connection network. The Network Burst divides a message into a small data chunk (called packet) and sends it in burst as much packet at a transfer opportunity. On the other hand, the Pipelined transfer propagates the packets to the devices in the network one after another. These techniques has been implemented on an FPGA as Maestro Link Controller with the Maestro Link Protocol that implements the two techniques above. According to the experimental evaluation with MLC, two techniques proposed above was effective for the optimization of inter-cluster communication.
Network interface of Maestro Cluster Network
(64bit PCI@66MHz, IEEE1394 200Mbps PHY, PowerPC603e, 64MB EDO DRAM are embeded)
To improve the optimization techniques for Maestro project, Maestro2 project proposed continuous network burst transfer and out-of-order switching mechanism. The continuous network burst transfer improved the Network Burst technique to keep longer burst transfer at a transfer opportunity. On the other hand, the out-of-order switching mechanism implemented a concurrent message transfer with a shred bus on the switch box. In addition, Maestro2 Cluster Network has been implemented with MLX (Maestro Link Protocol dupleX), and was evaluated with several communication experiments. The communication is made by a special message passing software called MMP. According to the evaluations, the continuous network burst transfer and the out-of-order switching were effective to improve the inter-cluster communication.
Network interface of Maestro2 Cluster Network
(64bit PCI@66MHz, LVDS 600Mbps PHY, PowerPC603e, 64MB SDRAM are embeded)