The arrival of 100K+ GPU clusters marks a new frontier in AI infrastructure. Standard communication stack meets new challenges as physical topologies span multiple datacenter buildings, introducing high bandwidth-delay product links where latency increases by up to 30× compared to intra-rack traffic. Furthermore, the transition toward Mixture-of-Experts architectures generating bursty all-to-all patterns that create transient congestion hotspots. These constraints, combined with an operational environment where hardware failures shift from anomalies to frequent occurrences, renders traditionally lightweight operations like initialization and resource management challenging.

Llama is a widely used open-source large language model.This paper presents the design and implementation of the parallelism techniques used in Llama 3 pre-training.To achieve efficient training on tens of thousands of GPUs, Llama 3 employs a combination of four-dimensional parallelism: fully sharded data parallelism, tensor parallelism, pipeline parallelism, and context parallelism.Beyond achieving efficiency through parallelism and model co-design, we

DLRM is a state-of-the-art recommendation system model that has gained widespread adoption across various industry applications. The large size of DLRM models, however, necessitates the use of multiple devices/GPUs for efficient training. A significant bottleneck in this process is the time-consuming all-to-all communication required to collect embedding data from all devices. To mitigate this, we introduce a method that employs error-bounded lossy compression to reduce the communication data size and accelerate DLRM training. We develop a novel error-bounded lossy compression algorithm, informed by an in-depth analysis of embedding data features, to achieve high compression ratios. Moreover, we introduce a dual-level adaptive strategy for error-bound adjustment, spanning both table-wise and iteration-wise aspects, to balance the compression benefits with the potential impacts on accuracy. We further optimize our compressor for PyTorch tensors on GPUs, minimizing compression overhead. Evaluation shows that our method achieves a 1.38 × training speedup with a minimal accuracy impact.

Deep learning recommendation models (DLRMs) have been used across many business-critical services at Meta and are the single largest AI application in terms of infrastructure demand in its data-centers. In this paper, we present Neo, a software-hardware co-designed system for high-performance distributed training of large-scale DLRMs. Neo employs a novel 4D parallelism strategy that combines table-wise, row-wise, column-wise, and data parallelism for training massive embedding operators in DLRMs. In addition, Neo enables extremely high-performance and memory-efficient embedding computations using a variety of critical systems optimizations, including hybrid kernel fusion, software-managed caching, and quality-preserving compression. Finally, Neo is paired with ZionEX, a new hardware platform co-designed with Neo's 4D parallelism for optimizing communications for large-scale DLRM training. Our evaluation on 128 GPUs using 16 ZionEX nodes shows that Neo outperforms existing systems by up to 40× for training 12-trillion-parameter DLRM models deployed in production.

In recent years, GPU-enhanced clusters have become more prevalent in High-Performance Computing (HPC), leading to a demand for more efficient multi-GPU communication. This makes it increasingly important to explore performance enhancements that can be attained through the communication middleware such as MPI, in order to fully take advantage of the GPUs available on these systems. In this paper, we propose locality-aware and adaptive schemes for hierarchical All-to-all collective communication on large-scale dense GPU systems. The proposed algorithms utilize the high bandwidth made available through the NVLink interconnect between GPUs in order to overlap communication latency. We focus on personalized and non-personalized all-to-all collective communication. These are components of modern scientific computing applications that utilize matrix transpose and three-dimensional Fast Fourier Transforms (FFT) and becoming more relevant for Deep Learning workloads with model and hybrid parallelisms. The performance evaluation with an application kernel performing three-dimensional FFT indicates that the proposed schemes for personalized all-to-all can lead to up to 15-25% lower execution time on 256 GPUs on the Lassen system. We demonstrate approximately 8% enhancement in training time for distributed K-FAC used in Deep Learning training on up to 128 GPUs. We also demonstrate approximately 22% and 30% improvement in the performance of non-personalized and personalized all-to-all benchmarks, respectively, compared to the state-of-the-art MPI libraries on the Summit and Lassen systems.

In the last decade, many scientific applications have been significantly accelerated by large-scale GPU systems. However, the movement of non-contiguous GPU-resident data is one of the most challenging components of scaling these applications using communication middleware like MPI. Although plenty of research has discussed improving noncontiguous data movement within communication middleware, the packing/unpacking operations on GPUs are still expensive. They cannot be hidden due to the limitation of MPI standard and the not-well-optimized designs in existing MPI implementations for GPU-resident data. Consequently, application developers tend to implement customized packing/unpacking kernels to improve GPU utilization by avoiding unnecessary synchronizations in MPI routines. However, this reduces productivity as well as performance as it cannot overlap the packing/unpacking operations with communication. In this paper, we propose a novel approach to achieve low-latency and high-bandwidth by dynamically fusing the packing/unpacking GPU kernels to reduce the expensive kernel launch overhead. The evaluation of the proposed designs shows up to 8X and 5X performance improvement for sparse and dense non-contiguous layout, respectively, compared to the state-of-the-art approaches on the Lassen system. Similarly, we observe up to 19X improvement over existing approaches on the ABCI system. Furthermore, the proposed design also outperforms the production libraries, such as SpectrumMPI, OpenMPI, and MVAPICH2, by many orders of magnitude.

The advanced fabrics like NVIDIA NVLink are enabling the deployment of dense Graphics Processing Unit (GPU) systems such as DGX-2 and Summit. With the wide adoption of large-scale GPU-enabled systems for distributed deep learning (DL) training, it is vital to design efficient communication such as the Allreduce operation to achieve near-ideal speedup at scale. In this paper, we propose a link-efficient scheme through NVLink-aware cooperative reduction kernels to significantly accelerate Allreduce operations for distributed deep learning applications. By overlapping computation and communication and maximizing utilization of all available NVLinks between CPU and GPU, as well as among GPUs, we demonstrate 1.8X performance improvement of Allreduce on 1,536 GPUs compared to state-of-the-art GPU-Aware MPI and NVIDIA NCCL libraries. Finally, we demonstrate 93.9% and 89.7% scaling efficiency (i.e., 15X and 172X speedup) for training ResNet-50 models using TensorFlow on a 16-GPU DGX-2 node and on 192-GPUs of the Summit system, respectively. To the best of our knowledge, this is the first study that achieves near-ideal scaling efficiency for distributed DL training and deals with designs tailored for cutting-edge systems like DGX-2 and Summit clusters.

The current wave of advances in Deep Learning (DL) have been triggered by the availability of large-scale datasets, efficient CPU and GPU hardware, and development of software frameworks like TensorFlow (TF). However, little exists in literature that addresses TensorFlow's distributed training capabilities. In this paper, we provide an in-depth performance characterization and design analysis for distributed TensorFlow. We present three key insights: 1) Horovod designs achieve better performance compared to the official gRPC-based approaches, 2) performance of Horovod design is heavily influenced by the time spent in gradient aggregation that uses the Allreduce primitive, and 3) performance of existing Horovod-MPI implementation is significantly worse compared to Horovod-NCCL. To address this limitation in Horovod-MPI, we propose a novel and efficient CUDA-Aware MPI Allreduce design that 1) exploits CUDA kernels to perform large reductions on the GPU, 2) uses a combination of bandwidth-optimal and latency-optimal algorithms, and 3) maintains a pointer cache to avoid CUDA-driver query overheads in the critical path. The proposed designs deliver 5×, 17×, and 29% better performance compared to NCCL2 for small, medium, and large messages. Our designs enable Horovod-MPI to beat state-of-the-art Horovod-NCCL2 by 3% and achieve 90% scaling efficiency for ResNet-50 training on 64 Pascal GPUs.

The recent advent of the NVLink interconnect and Peripheral Component Interconnect express (PCIe) switch has resulted in the creation of extremely dense Graphics Processing Unit (GPU) systems like Cray CS-Storm and NVIDIA DGX. In addition to the extremely high computational capability and communication capacity within a single machine, these systems expose novel capabilities like performing load-store operations from remote GPU memory across interconnects. While researchers have proposed solutions that take advantage of load-store semantics to provide support for high-performance datatype processing on CPUs, there exists no scholarly work on how one can orchestrate such high-performance datatype-based communication for GPU-resident data. In this paper, we take up this challenge and propose high-performance and architecture-aware designs for GPU-based non-contiguous datatype processing that uses the load-store semantics exposed by modern dense GPU systems. We demonstrate that the proposed solutions can reduce the overhead of datatype processing by up to 4.7X compared to the state-of-the-art schemes for GPU-based MILC communication kernel on an NVLink2-enabled dense GPU system. For a weather forecast application kernel, the proposed designs demonstrate up to 9.9X faster HaloExchange kernel among 64 GPUs over state-of-the-art designs. The proposed adaptive scheme also reports 10% higher throughput than existing designs for a 2D Jacobi solver on 16 GPUs. To the best of our knowledge, this is the first scholarly work that takes advantage of the zero-copy based load-store semantics to perform high-performance GPU to GPU derived datatype communication on modern dense GPU systems.

The recent advent of advanced fabrics like NVIDIA NVLink is enabling the deployment of dense Graphics Processing Unit (GPU) systems, e.g., DGX-2 and Summit. The Message Passing Interface (MPI) has been the dominant programming model to design distributed applications on such clusters. The MPI Tools Interface (MPI_T) provides an opportunity for performance tools and external software to introspect and understand MPI runtime behavior at a deeper level to detect performance and scalability issues. However, the lack of low-overhead and scalable monitoring tools have thus far prevented a comprehensive study of efficiency and utilization of high-performance interconnects such as NVLinks on high-performance GPU-enabled clusters. In this paper, we address this deficiency by proposing and designing an in-depth, real-time analysis, profiling, and visualization tool for high-performance GPU-enabled clusters with NVLinks. The proposed tool builds on the top of the OSU InfiniBand Network Analysis and Monitoring Tool (INAM). It provides insights into the efficiency of different communication patterns by examining the utilization of underlying GPU interconnects. The contributions of the proposed tool are two-fold: 1) domain scientists and system administrators can understand how applications and runtime libraries interact with underlying high-performance interconnects, and 2)Proposed tool enables designers of high-performance communication libraries to gain low-level knowledge to optimize existing designs and develop new algorithms to optimally utilize cutting-edge interconnects on GPU clusters. To the best of our knowledge, this is the first such tool which is capable of presenting a unified and holistic view of MPI-level and fabric level information for emerging NVLink-enabled high-performance GPU clusters.

In recent years, GPU-based platforms have received significant success for parallel applications. In addition to highly optimized computation kernels on GPUs, the cost of data movement on GPU clusters plays critical roles in delivering high performance for end applications. Many recent studies have been proposed to optimize the performance of GPUor CUDA-aware communication runtimes and these designs have been widely adopted in the emerging GPU-based applications. These studies mainly focus on improving the communication performance on native environments, i.e., physical machines, however GPU-based communication schemes on cloud environments are not well studied yet. This paper first investigates the performance characteristics of state-of-the-art GPU-based communication schemes on both native and container-based environments, which show a significant demand to design high-performance container-aware communication schemes in GPU-enabled runtimes to deliver near-native performance for end applications on clouds. Next, we propose the C-GDR approach to design high-performance Container-aware GPUDirect communication schemes on RDMA networks. C-GDR allows communication runtimes to successfully detect process locality, GPU residency, NUMA, architecture information, and communication pattern to enable intelligent and dynamic selection of the best communication and data movement schemes on GPU-enabled clouds. We have integrated C-GDR with the MVAPICH2 library. Our evaluations show that MVAPICH2 with C-GDR has clear performance benefits on container-based cloud environments, compared to default MVAPICH2-GDR and Open MPI. For instance, our proposed CGDR can outperform default MVAPICH2-GDR schemes by up to 66% on micro-benchmarks and up to 26% on HPC applications over a container-based environment.

Broadcast is a widely used operation in many streaming and deep learning applications to disseminate large amounts of data on emerging heterogeneous High-Performance Computing (HPC) systems. However, traditional broadcast schemes do not fully utilize hardware features for Graphics Processing Unit (GPU)-based applications. In this paper, a model-oriented analysis is presented to identify performance bottlenecks of existing broadcast schemes on GPU clusters. Next, streaming-based broadcast schemes are proposed to exploit InfiniBand hardware multicast (IB-MCAST) and NVIDIA GPUDirect technology for efficient message transmission. The proposed designs are evaluated in the context of using Message Passing Interface (MPI) based benchmarks and applications. The experimental results indicate improved scalability and up to 82 percent reduction of latency compared to the state-of-the-art solutions in the benchmark-level evaluation. Furthermore, compared to the state-of-the-art, the proposed design yields stable higher throughput for a synthetic streaming workload, and 1.3x faster training time for a deep learning framework.

Existing frameworks cannot train large DNNs that do not fit the GPU memory without explicit memory management schemes. In this paper, we propose OC-DNN - a novel Out-of-Core DNN training framework that exploits new Unified Memory features along with new hardware mechanisms in Pascal and Volta GPUs. OC-DNN has two major design components — 1) OC-Caffe; an enhanced version of Caffe that exploits innovative UM features like asynchronous prefetching, managed page-migration, exploitation of GPU-based page faults, and the cudaMemAdvise interface to enable efficient out-of-core training for very large DNNs, and 2) an interception library to transpar-ently leverage these cutting-edge features for other frameworks. We provide a comprehensive performance characterization of our designs. OC-Caffe provides comparable performance (to Caffe) for regular DNNs. OC-Caffe-Opt is up to 1.9X faster than OC-Caffe-Naive and up to 5X faster than optimized CPU-based training for out-of-core workloads. OC-Caffe also allows scale-up (DGX-1) and scale-out on multi-GPU clusters.

Traditionally, MPI runtimes have been designed for clusters with a large number of nodes. However, with the advent of MPI+CUDA applications and dense multi-GPU systems, it has become important to design efficient communication schemes. This coupled with new application workloads brought forward by Deep Learning frameworks like Caffe and Microsoft CNTK pose additional design constraints due to very large message communication of GPU buffers during the training phase. In this context, special-purpose libraries like NCCL have been proposed. In this paper, we propose a pipelined chain (ring) design for the MPI_Bcast collective operation along with an enhanced collective tuning framework in MVAPICH2-GDR that enables efficient intra-/internode multi-GPU communication. We present an in-depth performance landscape for the proposed MPI_Bcast schemes along with a comparative analysis of NCCL Broadcast and NCCL-based MPI_Bcast. The proposed designs for MVAPICH2-GDR enable up to 14X and 16.6X improvement, compared to NCCL-based solutions, for intra- and internode broadcast latency, respectively. In addition, the proposed designs provide up to 7% improvement over NCCL-based solutions for data parallel training of the VGG network on 128 GPUs using Microsoft CNTK. The proposed solutions outperform the recently introduced NCCL2 library for small and medium message sizes and offer comparable/better performance for very large message sizes.

Broadcast operations (e.g. MPI_Bcast) have been widely used in deep learning applications to exchange a large amount of data among multiple graphics processing units (GPUs). Recent studies have shown that leveraging the InfiniBand hardware-based multicast (IB-MCAST) protocol can enhance scalability of GPU-based broadcast operations. However, these initial designs with IB-MCAST are not optimized for multi-source broadcast operations with large messages, which is the common communication scenario for deep learning applications. In this paper, we first model existing broadcast schemes and analyze their performance bottlenecks on GPU clusters. Then, we propose a novel broadcast design based on message streaming to better exploit IB-MCAST and NVIDIA GPUDirect RDMA (GDR) technology for efficient large message transfer operation. The proposed design can provide high overlap among multi-source broadcast operations. Experimental results show up to 68% reduction of latency compared to state-of-the-art solutions in a benchmark-level evaluation. The proposed design also shows near-constant latency for a single broadcast operation as a system grows. Furthermore, it yields up to 24% performance improvement in the popular deep learning framework, Microsoft CNTK, which uses multi-source broadcast operations; notably, the performance gains are achieved without modifications to applications. Our model validation shows that the proposed analytical model and experimental results match within a 10% range. Our model also predicts that the proposed design outperforms existing schemes for multi-source broadcast scenarios with increasing numbers of broadcast sources in large-scale GPU clusters.

While GPUs are becoming common in HPC systems, the CPU is still responsible for managing both GPU-side and CPU-side compute, communication, and synchronization operations. For instance, if a result from a GPU-side computation is to be transferred to a remote destination, then the CPU must synchronize on GPU compute completion issuing a communication operation. Both CPU cycles and energy are consumed waiting for synchronization. In turn, this significantly affects overall application time and scalability (eg: strong scaling applications).In this work, we present techniques to decouple communication control flow between CPU and GPU on GPU-enabled systems with MPI+CUDA applications using the novel GPUDirect-aSync (GDS) mechanism. GDS allows the GPU to progress network communication with the goal of placing the CPU away from the critical path. To take advantage of GDS in MPI+CUDA applications, we introduce the notion of offloading MPI operations to CUDA streams (referred as MPI-GDS) which subsequently allow the GPU and the NIC to progress MPI communication in stream-order either before or after a CUDA operation. We also propose efficient designs/protocols to realize point-to-point communication operations that guarantee stream-ordering while achieving good performance. The proposed methods show good benefits with micro-benchmarks and up to 30% improvement in application-kernel pattern mimicking benchmark and up to 36% improvement with broadcast application-pattern simulation (in medium message range with 8 GPU nodes) in comparison with a pure MPI+CUDA application.

Accelerators like NVIDIA GPUs have changed the landscape of current HPC clusters to a great extent. Massive heterogeneous parallelism offered by these accelerators have led to GPU-Aware MPI libraries that are widely used for writing distributed parallel scientific applications. Compute-oriented collective operations like MPI_Reduce perform computation on data in addition to the usual communication performed by collectives. Historically, these collectives, due to their compute requirements have been implemented on CPU (or Host) only. However, with the advent of GPU technologies it has become important for MPI libraries to provide better design for their GPU (or Device) based versions.In this paper, we tackle the above challenges and provide designsand implementations for most commonly used compute-oriented collectives - MPI_Reduce, MPI_Allreduce, and MPI_Scan - for GPU clusters. We propose extensions to the state-of-the-art algorithms to fully take advantage of the GPU capabilities like GPUDirect RDMA (GDR) and CUDA compute kernel to efficiently perform these operations. With our new designs, we report reduced execution time for all computebased collectives up to 96 GPUs. Experimental results show an improvement of 50% for small messages and 85% for large messages using MPI_Reduce. For MPI_Allreduce and MPI_Scan, we report more than 40% reduction in time for large messages. Furthermore, analytical models are developed and evaluated to understand and predict the performance of proposed designs for extremely large-scale GPU clusters.

GPU accelerators are widely used in HPC clusters due to their massive parallelism and high throughput-per-watt. Data movement continues to be the major bottleneck on GPU clusters, more so when data is non-contiguous, which is common in scientific applications. CUDA-Aware MPI libraries optimize the non-contiguous data movement processing using latency oriented techniques such as using GPU kernels to accelerate the packing/unpacking operations. Although they optimize the latency of a single operation, the inherent restrictions of the designs limit their efficiency for throughput oriented patterns. Indeed, none of the existing designs fully exploit the massive parallelism of the GPUs to provide high throughput and efficient resources utilization by enabling maximal overlap. In this paper, we propose novel designs for CUDA-Aware MPI libraries to achieve efficient GPU resource utilization and maximal overlap between CPUs and GPUs for non-contiguous data processing and movement. The proposed designs take advantage of several CUDA features, such as Hyper-Q/multi-streams and callback function, to deliver high performance and efficiency. To the best of our knowledge, this is the first such study to provide high throughput and efficient resource utilization for non-contiguous MPI data processing and movement to/from GPUs. The performance evaluation with the proposed designs using DDTBench shows up to 54%, 67%, 61% performance improvement on the SPECFEM3D_oc, SPECFEM3D_cm and WRF_y_sa benchmarks respectively for intra-node inter-GPU ping-pong experiments. The proposed designs also deliver up to 33% improvement on the total execution time over the existing designs for the HaloExchange-based application kernel that models the communication pattern of the MeteoSwiss weather forecasting model over 32 GPU nodes on Wilkes GPU cluster.

High-performance streaming applications are beginning to leverage the compute power offered by graphics processing units (GPUs) and high network throughput offered by high performance interconnects such as InfiniBand (IB) to boost their performance and scalability. These applications rely heavily on broadcast operations to move data, which is stored in the host memory, from a single source-typically live-to multiple GPU-based computing sites. While homogeneous broadcast designs take advantage of IB hardware multicast feature to boost their performance, their heterogeneous counterpart requires an explicit data movement between Host and GPU, which significantly hurts the overall performance. There is a dearth of efficient heterogeneous broadcast designs for streaming applications especially on emerging multi-GPU configurations. In this work, we propose novel techniques to fully and conjointly take advantage of NVIDIA GPUDirect RDMA (GDR), CUDA inter-process communication (IPC) and IB hardware multicast features to design high-performance heterogeneous broadcast operations for modern multi-GPU systems. We propose intra-node, topology-aware schemes to maximize the performance benefits while minimizing the utilization of valuable PCIe resources. Further, we optimize the communication pipeline by overlapping the GDR + IB hardware multicast operations with CUDA IPC operations. Compared to existing solutions, our designs show up to 3X improvement in the latency of a heterogeneous broadcast operation. Our designs also show up to 67% improvement in execution time of a streaming benchmark on a GPU-dense Cray CS-Storm system with 88 GPUs.

GPUDirect RDMA (GDR) brings the high-performance communication capabilities of RDMA networks like InfiniBand (IB) to GPUs (referred to as "Device"). It enables IB network adapters to directly write/read data to/from GPU memory. Partitioned Global Address Space (PGAS) programming models, such as OpenSHMEM, provide an attractive approach for developing scientific applications with irregular communication characteristics by providing shared memory address space abstractions, along with one-sided communication semantics. However, current approaches and designs of OpenSHMEM on GPU clusters do not take advantage of the GDR features leading to inefficiencies and sub-optimal performance. In this paper, we analyze the performance of various OpenSHMEM operations with different inter-node and intra-node communication configurations (Host-to-Device, Device-to-Device, and Device-to-Host) on GPU based systems. We propose novel designs that ensure "truly one-sided" communication for the different inter-/intra-node configurations identified above while working around the hardware limitations. To the best of our knowledge, this is the first work that investigates GDR-aware designs for OpenSHMEM communication operations. Experimental evaluations indicate 2.5X and 7X improvement in point-point communication for intra-node and inter-node, respectively. The proposed framework achieves 2.2µs for an intra-node 8 byte put operation from Host-to-Device, and 3.13µs for an inter-node 8 byte put operation from GPU to remote GPU. With Stencil2D application kernel from SHOC benchmark suite, we observe a 19% reduction in execution time on 64 GPU nodes. Further, for GPULBM application, we are able to improve the performance of the evolution phase by 53% and 45% on 32 and 64 GPU nodes, respectively.

Long-distance Wi-Fi technology has shown promise in several network applications that cannot utilize conventional technologies effectively. Although the network performance of Wi-Fi technology is affected by a number of environmental factors, there is a dearth of long-term, continuous, and systematic studies on the correlations between those factors and the technology's performance. In this study, we deployed a long-distance Wi-Fi testbed on our campus and conducted a one-year experiment. Comprehensive data analysis of the measurement results shows that rainfall is the major weather attribute that affects the network performance of long-distance Wi-Fi links. In addition, the performance is highly correlated to human activities in the immediate vicinity. The results also demonstrate it is possible to infer people's daily routines on campus by exploiting the long-term measurement data.