InfiniBand

Abstract: Long training times for high-accuracy deep neural networks (DNNs) impede research into new DNN architectures and slow the development of high-accuracy DNNs. In this paper we present FireCaffe, which successfully scales deep neural network training across a cluster of GPUs. We also present a number of best practices to aid in comparing advancements in methods for scaling and accelerating the training of deep neural networks. The speed and scalability of distributed algorithms is almost always limited by the overhead of communicating between servers, DNN training is not an exception to this rule. Therefore, the key consideration here is to reduce communication overhead wherever possible, while not degrading the accuracy of the DNN models that we train. Our approach has three key pillars. First, we select network hardware that achieves high bandwidth between GPU servers – Infiniband or Cray interconnects are ideal for this. Second, we consider a number of communication algorithms, and we find that reduction trees are more efficient and scalable than the traditional parameter server approach. Third, we optionally increase the batch size to reduce the total quantity of communication during DNN training, and we identify hyperparameters that allow us to reproduce the small-batch accuracy while training with large batch sizes. When training GoogLeNet and Network-in-Network on ImageNet, we achieve a 47x and 39x speedup, respectively, when training on a cluster of 128 GPUs.

Abstract: Currently, I/O device virtualization models in virtual machine (VM) environments require involvement of a virtual machine monitor (VMM) and/or a privileged VM for each I/O operation, which may turn out to be a performance bottleneck for systems with high I/O demands, especially those equipped with modern high speed interconnects such as InfiniBand. In this paper, we propose a new device virtualization model called VMM-bypass I/O, which extends the idea of OS-bypass originated from user-level communication. Essentially, VMM-bypass allows time-critical I/O operations to be carried out directly in guest VMs without involvement of the VMM and/or a privileged VM. By exploiting the intelligence found in modern high speed network interfaces, VMM-bypass can significantly improve I/O and communication performance for VMs without sacrificing safety or isolation. To demonstrate the idea of VMM-bypass, we have developed a prototype called Xen-IB, which offers InfiniBand virtualization support in the Xen 3.0 VM environment. Xen-IB runs with current InfiniBand hardware and does not require modifications to existing user-level applications or kernel-level drivers that use InfiniBand. Our performance measurements show that Xen-IB is able to achieve nearly the same raw performance as the original InfiniBand driver running in a non-virtualized environment.

Abstract: JUWELS is a multi-petaflop modular supercomputer operated by Julich Supercomputing Centre at Forschungszentrum Julich as a European and national supercomputing resource for the Gauss Centre for Supercomputing. In addition, JUWELS serves the Earth system modeling community within the Helmholtz Association. The first module deployed in 2018, is a Cluster module based on the BullSequana X1000 architecture with Intel Xeon Skylake-SP processors and Mellanox EDR InfiniBand. An extension by a second Booster module is scheduled for deployment in 2020.

Abstract: Long training times for high-accuracy deep neural networks (DNNs) impede research into new DNN architectures and slow the development of high-accuracy DNNs. In this paper we present FireCaffe, which successfully scales deep neural network training across a cluster of GPUs. We also present a number of best practices to aid in comparing advancements in methods for scaling and accelerating the training of deep neural networks. The speed and scalability of distributed algorithms is almost always limited by the overhead of communicating between servers; DNN training is not an exception to this rule. Therefore, the key consideration here is to reduce communication overhead wherever possible, while not degrading the accuracy of the DNN models that we train. Our approach has three key pillars. First, we select network hardware that achieves high bandwidth between GPU servers -- Infiniband or Cray interconnects are ideal for this. Second, we consider a number of communication algorithms, and we find that reduction trees are more efficient and scalable than the traditional parameter server approach. Third, we optionally increase the batch size to reduce the total quantity of communication during DNN training, and we identify hyperparameters that allow us to reproduce the small-batch accuracy while training with large batch sizes. When training GoogLeNet and Network-in-Network on ImageNet, we achieve a 47x and 39x speedup, respectively, when training on a cluster of 128 GPUs.