A Novel Approach to Supporting Communicators for In-Switch Processing of MPI Collectives
A Novel Approach to Supporting Communicators for In-Switch Processing of MPI Collectives
复制标题
支持通信器进行 MPI 集合的交换内处理的新方法
DOI：
--
复制
发表时间：
2019
期刊：
影响因子：
--
通讯作者：
A. Skjellum
中科院分区：
文献类型：
--
作者：
Joshua Stern;Qingqing Xiong;A. Skjellum
MPI集体操作通常是HPC应用程序中的性能杀手;我们试图通过将它们卸载到交换机本身的硬件来解决这个瓶颈。我们已经从以前的工作中看到，将集体移动到网络中提供了显着的性能优势。然而，在为次级传播者集体提供支持方面进展甚微。我们引入了一种新的机制，使硬件支持大量的通信器的任意形状，可扩展到非常大的系统。我们已经将此支持集成到交换机内硬件加速器中，以实现对MPI通信器的支持和MPI集合的完全卸载。虽然这种机制是普遍适用的，我们在FPGA集群中实现它; FPGA提供了耦合通信和计算的能力，因此提供了一个理想的测试平台。初步结果表明，我们可以在可接受的硬件成本，包括一个10倍的加速比传统集群的短消息集体在不规则的内部通信器的性能大幅提高。
MPI collective operations can often be performance killers in HPC applications; we seek to solve this bottleneck by offloading them to hardware within the switch itself. We have seen from previous work that moving collectives into the network offers significant performance benefits. However, there has been little advancement in providing support for sub-communicator collectives. We introduce a novel mechanism that enables the hardware to support a large number of communicators of arbitrary shape that is scalable to very large systems. We have integrated this support into an in-switch hardware accelerator to implement support for MPI communicators and full offload of MPI collectives. While this mechanism is universally applicable, we implement it in an FPGA cluster; FPGAs provide the ability to couple communication and computation and so provide an ideal testbed. Preliminary results show that we can achieve substantial performance improvement at acceptable hardware cost, including a 10× speedup over conventional clusters for short message collectives over irregular intra-communicators.