Evaluating Modern GPU Interconnect: PCIe, NVLink, NV-SLI, NVSwitch and GPUDirect (1903.04611v1)

Abstract: High performance multi-GPU computing becomes an inevitable trend due to the ever-increasing demand on computation capability in emerging domains such as deep learning, big data and planet-scale simulations. However, the lack of deep understanding on how modern GPUs can be connected and the real impact of state-of-the-art interconnect technology on multi-GPU application performance become a hurdle. In this paper, we fill the gap by conducting a thorough evaluation on five latest types of modern GPU interconnects: PCIe, NVLink-V1, NVLink-V2, NVLink-SLI and NVSwitch, from six high-end servers and HPC platforms: NVIDIA P100-DGX-1, V100-DGX-1, DGX-2, OLCF's SummitDev and Summit supercomputers, as well as an SLI-linked system with two NVIDIA Turing RTX-2080 GPUs. Based on the empirical evaluation, we have observed four new types of GPU communication network NUMA effects: three are triggered by NVLink's topology, connectivity and routing, while one is caused by PCIe chipset design issue. These observations indicate that, for an application running in a multi-GPU node, choosing the right GPU combination can impose considerable impact on GPU communication efficiency, as well as the application's overall performance. Our evaluation can be leveraged in building practical multi-GPU performance models, which are vital for GPU task allocation, scheduling and migration in a shared environment (e.g., AI cloud and HPC centers), as well as communication-oriented performance tuning.

• The paper presents an empirical evaluation of various GPU interconnects, detailing their impact on latency and bandwidth in multi-GPU setups.
• The paper uses comprehensive microbenchmarking to reveal significant NUMA effects and performance enhancements using NVLink and NVSwitch.
• The paper underscores the need for advanced multi-GPU programming models to fully leverage modern interconnects in HPC and deep learning applications.

Evaluating Modern GPU Interconnects: PCIe, NVLink, NV-SLI, NVSwitch, and GPUDirect

Introduction

The proliferation of complex multi-GPU setups has been driven by the constantly growing computational needs in areas like deep learning and high-performance computing (HPC). This paper thoroughly investigates the characteristics and performance implications of several cutting-edge GPU interconnect technologies, including PCIe, NVLink-V1/V2, NV-SLI, and NVSwitch, as well as GPUDirect-enabled InfiniBand. The goal is to provide empirical insights that inform optimal GPU communication configurations, enhance application performance models, and support multi-GPU execution frameworks.

Modern GPU Interconnect Technologies

The investigation covers several types of GPU interconnects:

• PCIe: A traditional high-speed serial computing expansion bus, typically a bottleneck in GPU-accelerated systems due to its relatively slow speed compared to the interconnects like NVLink.

  Figure 1: PCIe and NVLink-V1/V2 topology for P100-DGX-1 and V100-DGX-1.

• NVLink: NVLink-V1 and NVLink-V2 significantly improve bandwidth for GPU-to-GPU communications via high-speed signaling, reducing the dependency on slower PCIe connections. NVLink-V2 further enhances this with increased bandwidth and additional link-slots per GPU.

  Figure 2: NVLink interconnect topology for SummitDev and Summit.

• NV-SLI: Originally a solution for graphics applications, the newer NV-SLI modifies NVLink for multi-GPU co-rendering and computing, offering high-bandwidth interconnect between paired GPUs.
• NVSwitch: Designed for the DGX-2 systems, NVSwitch facilitates high-throughput, all-to-all GPU communication within a single node, expanding upon NVLink capabilities to support larger clusters of interconnected GPUs.

Figure 3: NVSwitch interconnect topology in DGX-2.

• GPUDirect: By enabling direct memory access between GPUs and peripherals like InfiniBand adapters, GPUDirect reduces latency and increases bandwidth efficiency for inter-node GPU communications.

GPU Interconnect Microbenchmarking

Comprehensive microbenchmarking was conducted to analyze latency, bandwidth, and NUMA effects across various intranode and internode communication patterns:

• Intra-node P2P Communication: Examined via PCIe, NVLink, NV-SLI, and NVSwitch, highlighting NUMA effects that arise from topology, connectivity, and routing. NVLink and NVSwitch generally demonstrated superior bandwidth capabilities compared to PCIe.

Figure 4: P100/V100-DGX-1 P2P communication latency indicative of NUMA effects.

• Intra-node CL Communication: Implementations leveraging NCCL showcased improvements in communication efficiency. Understanding the topology and integrating optimally structured communication patterns (e.g., ring networks) could mitigate potential bottlenecks.
• Inter-node P2P and CL Communication: Leveraging GPUDirect significantly improved bandwidth efficiency, contrasting scenarios with traditional memory access methods. Applications such as all-reduce could particularly benefit from the superior throughput of GPUDirect-RDMA.

Benchmarking and Observations

The paper utilized the Tartan Benchmark Suite to analyze performance implications of these interconnect technologies on real-world multi-GPU applications, focusing on both scale-up and scale-out configurations:

• Intra-node Scale-up: Without optimizing for inter-GPU communication, strong scaling benefits from NVLink enhancements were limited. Further gains require a paradigm shift in parallelization models to capitalize on high-speed interconnects.

  Figure 5: Normalized latency reduction by NVLink-V1 and NCCL-V2 of weak scaling for single-node scaling-up on NVIDIA P100-DGX-1.

• Inter-node Scale-out: Applications exhibited notable performance improvements when leveraging GPUDirect-RDMA, underscoring the necessity for efficient inter-node communication strategies. Faster interconnects, such as InfiniBand, present clear advantages in multi-node HPC settings.

  Figure 6: Performance speedup by InfiniBand GPUDirect-RDMA of strong scaling for multi-node scaling-out on ORNL SummitDev.

Conclusion

This paper systematically evaluates the latest GPU interconnect technologies, elucidating their effects on HPC and machine learning applications. The distinct NUMA effects observed, along with the heterogeneous nature of these interconnect systems, highlight the need for robust multi-GPU programming models that can dynamically leverage these high-speed communication channels. Future developments should focus on refining multi-GPU execution frameworks and performance models to maximize the potential of modern GPU interconnects in computing environments.