Distributed training of large deep neural networks requires frequent exchange of massive data between machines, thus communication efficiency is a major concern. Existing compressed communication methods are either not compatible with large batch optimization algorithms, or do not provide sufficient speedup in large scale. In this paper, we combine sparsification-based gradient compression with the layer-wise adaptive moments optimizer for large batch training (LAMB). We propose SLAMB, a novel communication-efficient optimizer that supports large batch sizes and scales to thousands of GPUs. SLAMB employs momentum masking, local error compensation, and element-wise adaptive rescaling to achieve accurate layer-wise weight updates, which translates to fast convergence for very large batches. Our empirical results show that, compared to the state-of-the-art, SLAMB transmits half the amount of data in large-batch BERT pre-training, without sacrificing accuracy. Moreover, SLAMB achieves excellent scalability in large computing infrastructures. For instance, SLAMB with 128 GPUs reduces the training time of Swin Transformer pre-training on ImageNet to 5.35 hours, which is 2 hours faster than the state-of-the-art. At the extreme, we trained BERT-XL (2.8B parameters) on 1,024 NVIDIA A100 GPUs, where SLAMB achieved 90% scaling efficiency.

State-of-the-art systolic array-based accelerators adopt the traditional im2col algorithm to accelerate the inference of convolutional layers. However, traditional im2col cannot efficiently support AI backpropagation. Backpropagation in convolutional layers involves performing transposed convolution and dilated convolution, which usually introduces plenty of zero-spaces into the feature map or kernel. The zero-space data reorganization interfere with the continuity of training and incur additional and non-negligible overhead in terms of off- and on-chip storage, access and performance. Since countermeasures for backpropagation are rarely proposed, we propose BP-im2col, a novel im2col algorithm for AI backpropagation, and implement it in RTL on a TPU-like accelerator. Experiments on TPU-like accelerator indicate that BP-im2col reduces the backpropagation runtime by 34.9% on average, and reduces the bandwidth of off-chip memory and on-chip buffers by at least 22.7% and 70.6% respectively, over a baseline accelerator adopting the traditional im2col. It further reduces the additional storage overhead in the backpropagation process by at least 74.78%.

Generalized Sparse Matrix-Matrix Multiplication (SpGEMM) is a critical kernel in domains like graph analytic and scientific computation. As a kind of classical special-purpose architecture, systolic arrays were first used for complex computing problems, e.g., matrix multiplication. However, classical systolic arrays are not efficient enough when handling sparse matrices due to the fact that the PEs containing zero-valued entries perform unnecessary operations that do not contribute to the result. Accordingly, in this paper, we propose Mentha, a framework that enables systolic arrays to accelerate sparse matrix computation by employing a sparse-packing algorithm suitable for various dataflow of systolic array. Firstly, Mentha supports both online and offline methods. By packing the rows or columns of the sparse matrix, the zero-valued items in the matrix are significantly reduced and the density of the matrix is improved. In addition, acceleration benefits can be obtained by the adaptation scheme even with limited resources. Moreover, we reconfigure PEs in systolic arrays at a low cost (1.28x in area, 1.21x in power) and find that our method outperforms TPU-like systolic arrays by 1.2~3.3x in terms of SpMM and 1.3~4.4x in terms of SpGEMM when dealing with moderately sparse matrices (sparsity < 0.9), while its performance is at least 9.7x better than cuSPARSE. Furthermore, experimental results show a FLOPs reduction of roughly 3.4x in the neural network.

The systolic array provides extremely high efficiency for running matrix multiplication, and is one of the mainstream architectures of today’s deep learning accelerators. In order to develop efficient accelerators, people usually employ simulators to make design trade-offs. However, current simulators suffer from coarse-grained modeling methods and ideal assumptions, which limits their ability of describing structural characteristics of systolic arrays. In addition, they do not support the exploration of microarchitecture. This paper presents TILE-SIM, a computing-centric systematic method for evaluating systolic array accelerators by using an event-driven method. TILE-SIM can obtain accurate results and provide the best mapping scheme for different workload due to its fine-grained modeling technique and deny of ideal assumption. Experimental results show that TILE-SIM plays a significant role in design trade-offs and outperforms state-of-the-art simulators, with an accuracy of more than 95%.

The simulation of turbulent flow, the most common form of fluid, is indispensable in computational fluid dynamics (CFD). The synthetic eddy method (SEM) generates the turbulent inflow and is adopted as the inlet boundary condition of simulation. However, SEM is time-consuming and can significantly slow down the simulation process which occupies 58% of the whole computational time. This is highly inefficient especially since SEM is only used as the inlet. In this paper we propose an efficient alternative. In particular, we leverage CNN+LSTM to replace SEM to obtain the turbulence statistics and combine it with link-wise artificial compressibility method (LW-ACM), which is a fast numerical method of CFD. We validate the predicted results by CNN+LSTM and prove that our model can provide the correct turbulence statistics even after a long time. Experiment results show that our CNN+LSTM module achieves over 15 × speedup compared with SEM, which greatly reduces the time consumption of turbulent inflow generation (from 58% to 7%). As a result, the whole time of turbulent flow simulation is more than halved. Compared with a newly released GPU-accelerated standard lattice Boltzmann method solver, our combination of CNN+LSTM and LW-ACM is about 8.6 × faster. Among all studies reported to date, our work is the fastest implementation for simulating turbulent channel flow, an important step for the field of fast CFD analysis.

Efficient collective communication is crucial to parallel-computing applications such as distributed training of large-scale recommendation systems and natural language processing models. Existing collective communication libraries focus on optimizing operations for dense inputs, resulting in transmissions of many zeros when inputs are sparse. This counters current trends that see increasing data sparsity in large models.

Today's cloud-based online services are underpinned by distributed key-value stores (KVSs). Keys and values are distributed across back-end servers in such scale-out systems. One primary real-life performance bottleneck occurs when storage servers suffer from load imbalance under skewed workloads. In this paper, we present KVSwitch, a centralized self-managing load balancer that leverages the power and flexibility of emerging programmable switches. The balance is achieved through dynamically predicting the hot items and creating replication strategies according to KVS loading. To overcome the challenges in realizing KVSwitch given the limitations of the switch hardware, we decompose KVSwitch's functions and carefully design them for the heterogeneous processors inside the switch. We prototype KVSwitch in a Tofino switch. Experimental results show that our solution can effectively keep the KVS servers balanced even under highly skewed workloads. Furthermore, KVSwitch only replicates 70% of hot items and consumes 9.88% of server memory rather than simply replicating all hot items to each server.