面向高性能计算应用的双总线型聚合物波导阵列研究

In the recent research of High-Performance Computing(HPC), interconnection network becomes the main barrier to improve computing ability.Due to this, an innovative bi-directional optical bus architecture based on multimode polymer waveguides was proposed to solve the problem of high speed interconnection..The whole architecture consists of two parallel optical bus waveguides, which can transmit optical signals along two opposite directions. The optical signals from the transmitter (T) of each transceiver will be split into two beams and transmitted bi-directionally into the two bus waveguides through two unidirectional branch waveguides. The receiver (R) of each transceiver is also capable of receiving optical signals from both of the two bus waveguides, benefited from the two unidirectional branch waveguides that are connected with them. The two parallel optical bus waveguides in conjunction with the unidirectional branch waveguides ensure completely non-blocking interconnection among any existing transceiver without any wiring crossing. . Superior to any point-to-point optical waveguide, which is useful to most high performance computer (HPC) system only when it is combined with electrical switch backplanes, the implemented optical bus structure is capable of broadcasting and receiving high speed data at 10Gbit/sec among multiple points as a high performance optical backplane or interchip optical intercontects. In addition,this polymer waveguide array have also the advantages of enhanced bandwidth, increased reliability, package compatibility and significantly lower fabrication cost.

在高性能计算系统研究中，互连网络成为制约系统性能提高的关键。基于此，研究提出了一种面向高性能计算应用的双总线型聚合物多模波导阵列方案，整个结构由总线波导、分支波导和Y分支组成，通过优化设计总线波导和分支波导的尺寸比例和分支波导的弯曲半径，来实现多节点间高速光互连。.该方案借鉴了传统冯.诺依曼总线架构计算机模型，完全不同于基于MESH或Crossbar的光互连网络；它既克服了光电集成互连网络存在电交换互连瓶颈的问题，又改进了点到点光互连技术难以有效实现数据广播分发和接收的不足，且该波导阵列具有可扩展，制作工艺成熟和成本较低等优点，可应用于高速光学背板或PCB板上芯片间高速光互连，是对未来高性能计算机或超级计算机系统架构的一种有效尝试和探索。

在高性能计算系统研究中，互连网络成为制约系统性能提高的关键。项目提出一种新的共享总线型聚合物波导阵列设计方案；它采用多模光波导作为光互连层，通过光束追迹法来优化设计总线波导和分支波导，实现总线波导上的光信号能等比例分配到各端口；再运用基于二阶pade算子的光束传输法（BPM）来仿真分析多模光信号传输与光强分布变化的关系，实现光信号有效上载到光总线。研究发现：分支波导分别采用38微米、40微米和50微米，四端口输出功率分别为：0.232、0.262、0.223和0.241。同时建立波导弯曲半径与传输损耗以及波导交叉与信号串扰的关系模型，以获得最低插入损耗光传输通道；研究发现：宽度为50μm的多模波导其90°弯曲半径最佳值是13.5μm；且波导宽度越大，最佳90°半径弯曲损耗越大。最后进行波导阵列的实际制作与加工，波导包层和芯层分别采用OE-4141和OE-4140，直波导光传输损耗在850nm时低至0.05dB/cm；通过聚合物波导阵列实际测试发现，波导端面的平整度以及分支波导与总线波导的粘连度成为产生波导传输损耗的主要原因。.项目设计并验证了总线结构聚合物波导阵列在光互连网络应用的可行性，证明制作工艺成为制约聚合物波导实际应用的关键。该波导阵列借鉴了传统冯.诺依曼总线架构计算机模型，它完全不同于基于MESH或Crossbar的光互连网络，又充分发挥了聚合物波导高速光互连技术的优势，是对发展高性能计算系统的一种有效探索和尝试。