Search

Modern accelerators like GPUs are increasingly executing independent operations concurrently to improve the device's compute utilization. However, effectively harnessing it on GPUs for important primitives such as general matrix multiplications (GEMMs) remains challenging. Although modern GPUs have significant hardware and software support for GEMMs, their kernel implementations and optimizations typically assume each kernel executes in isolation and can utilize all GPU resources. This approach is highly efficient when kernels execute in isolation, but causes significant resource contention and slowdowns when kernels execute concurrently. Moreover, current approaches often only statically expose and control parallelism within an application, without considering runtime information such as varying input size and concurrent applications -- often exacerbating contention. These issues limit performance benefits from concurrently executing independent operations. Accordingly, we propose GOLDYLOC, which considers the global resources across all concurrent operations to identify performant GEMM kernels, which we call globally optimized (GO)-Kernels. Moreover, GOLDYLOC introduces a lightweight dynamic logic which considers the dynamic execution environment for available parallelism and input sizes to execute performant combinations of concurrent GEMMs on the GPU. Overall, GOLDYLOC improves performance of concurrent GEMMs on a real GPU by up to 2$\times$ (18% geomean per workload) and provides up to 2.5$\times$ (43% geomean per workload) speedups over sequential execution.

Large-scale computing systems are increasingly using accelerators such as GPUs to enable peta- and exa-scale levels of compute to meet the needs of Machine Learning (ML) and scientific computing applications. Given the widespread and growing use of ML, including in some scientific applications, optimizing these clusters for ML workloads is particularly important. However, recent work has demonstrated that accelerators in these clusters can suffer from performance variability and this variability can lead to resource under-utilization and load imbalance. In this work we focus on how clusters schedulers, which are used to share accelerator-rich clusters across many concurrent ML jobs, can embrace performance variability to mitigate its effects. Our key insight to address this challenge is to characterize which applications are more likely to suffer from performance variability and take that into account while placing jobs on the cluster. We design a novel cluster scheduler, PAL, which uses performance variability measurements and application-specific profiles to improve job performance and resource utilization. PAL also balances performance variability with locality to ensure jobs are spread across as few nodes as possible. Overall, PAL significantly improves GPU-rich cluster scheduling: across traces for six ML workload applications spanning image, language, and vision models with a variety of variability profiles, PAL improves geomean job completion time by 42%, cluster utilization by 28%, and makespan by 47% over existing state-of-the-art schedulers.

Large Language Models increasingly rely on distributed techniques for their training and inference. These techniques require communication across devices which can reduce scaling efficiency as the number of devices increases. While some distributed techniques can overlap, and thus, hide this communication with independent computations, techniques such as Tensor Parallelism (TP) inherently serialize communication with model execution. One approach to hide this serialized communication is to interleave it with the producer operation (of the communicated data) in a fine-grained manner. However, this fine-grained interleaving of communication and computation in software can be difficult. Furthermore, as with any concurrent execution, it requires compute and memory resources to be shared between computation and communication, causing resource contention that reduces overlapping efficacy. To overcome these challenges, we propose T3 which applies hardware-software co-design to transparently overlap serialized communication while minimizing resource contention with compute. T3 transparently fuses producer operations with the subsequent communication via a simple configuration of the producer's output address space and requires minor software changes. At the hardware level, T3 adds a lightweight track and trigger mechanism to orchestrate the producer's compute, and communication. It further uses compute-enhanced memories for communication's attendant compute. As a result, T3 reduces resource contention, and efficiently overlaps serialized communication with computation. For important Transformer models like T-NLG, T3 speeds up communication-heavy sublayers by 30% geomean (max 47%) and reduces data movement by 22% geomean (max 36%). Furthermore, T3's benefits persist as models scale: geomean 29% for sublayers in $\sim$500-billion parameter models, PALM and MT-NLG., Comment: To appear at the International Conference on Architectural Support for Programming Languages and Operating Systems (ASPLOS) 2024

Scaling neural network models has delivered dramatic quality gains across ML problems. However, this scaling has increased the reliance on efficient distributed training techniques. Accordingly, as with other distributed computing scenarios, it is important to understand how will compute and communication scale relative to one another as models scale and hardware evolves? A careful study which answers this question can better guide the design of future systems which can efficiently train future large models. Accordingly, this work provides a comprehensive multi-axial (algorithmic, empirical, hardware evolution) analysis of compute vs. communication (Comp-vs.-Comm) scaling for future Transformer models on future hardware. First, our algorithmic analysis shows that compute generally enjoys an edge over communication as models scale. However, since memory capacity scales slower than compute, these trends are being stressed. Next, we quantify this edge by empirically studying how Comp-vs.-Comm scales for future models on future hardware. To avoid profiling numerous Transformer models across many setups, we extract execution regions and project costs using operator models. This allows a spectrum (hundreds) of future model/hardware scenarios to be accurately studied ($<$15% error), and reduces profiling costs by 2100$\times$. Our experiments show that communication will be a significant portion (40-75%) of runtime as models and hardware evolve. Moreover, communication which is hidden by overlapped computation in today's models often cannot be hidden in future, larger models. Overall, this work highlights the increasingly large role communication will play as models scale and discusses techniques and upcoming technologies that can help address it.

Scientists are increasingly exploring and utilizing the massive parallelism of general-purpose accelerators such as GPUs for scientific breakthroughs. As a result, datacenters, hyperscalers, national computing centers, and supercomputers have procured hardware to support this evolving application paradigm. These systems contain hundreds to tens of thousands of accelerators, enabling peta- and exa-scale levels of compute for scientific workloads. Recent work demonstrated that power management (PM) can impact application performance in CPU-based HPC systems, even when machines have the same architecture and SKU (stock keeping unit). This variation occurs due to manufacturing variability and the chip's PM. However, while modern HPC systems widely employ accelerators such as GPUs, it is unclear how much this variability affects applications. Accordingly, we seek to characterize the extent of variation due to GPU PM in modern HPC and supercomputing systems. We study a variety of applications that stress different GPU components on five large-scale computing centers with modern GPUs: Oak Ridge's Summit, Sandia's Vortex, TACC's Frontera and Longhorn, and Livermore's Corona. These clusters use a variety of cooling methods and GPU vendors. In total, we collect over 18,800 hours of data across more than 90% of the GPUs in these clusters. Regardless of the application, cluster, GPU vendor, and cooling method, our results show significant variation: 8% (max 22%) average performance variation even though the GPU architecture and vendor SKU are identical within each cluster, with outliers up to 1.5X slower than the median GPU. These results highlight the difficulty in efficiently using existing GPU clusters for modern HPC and scientific workloads, and the need to embrace variability in future accelerator-based systems., Comment: 14 pages, 18 figures, to appear at The 34th International Conference for High Performance Computing, Networking, Storage, and Analysis (SC '22)

Transfer learning in natural language processing (NLP), as realized using models like BERT (Bi-directional Encoder Representation from Transformer), has significantly improved language representation with models that can tackle challenging language problems. Consequently, these applications are driving the requirements of future systems. Thus, we focus on BERT, one of the most popular NLP transfer learning algorithms, to identify how its algorithmic behavior can guide future accelerator design. To this end, we carefully profile BERT training and identify key algorithmic behaviors which are worthy of attention in accelerator design. We observe that while computations which manifest as matrix multiplication dominate BERT's overall runtime, as in many convolutional neural networks, memory-intensive computations also feature prominently. We characterize these computations, which have received little attention so far. Further, we also identify heterogeneity in compute-intensive BERT computations and discuss software and possible hardware mechanisms to further optimize these computations. Finally, we discuss implications of these behaviors as networks get larger and use distributed training environments, and how techniques such as micro-batching and mixed-precision training scale. Overall, our analysis identifies holistic solutions to optimize systems for BERT-like models.

The ubiquity of deep neural networks (DNNs) continues to rise, making them a crucial application class for hardware optimizations. However, detailed profiling and characterization of DNN training remains difficult as these applications often run for hours to days on real hardware. Prior works exploit the iterative nature of DNNs to profile a few training iterations. While such a strategy is sound for networks like convolutional neural networks (CNNs), where the nature of the computation is largely input independent, we observe in this work that this approach is sub-optimal for sequence-based neural networks (SQNNs) such as recurrent neural networks (RNNs). The amount and nature of computations in SQNNs can vary for each input, resulting in heterogeneity across iterations. Thus, arbitrarily selecting a few iterations is insufficient to accurately summarize the behavior of the entire training run. To tackle this challenge, we carefully study the factors that impact SQNN training iterations and identify input sequence length as the key determining factor for variations across iterations. We then use this observation to characterize all iterations of an SQNN training run (requiring no profiling or simulation of the application) and select representative iterations, which we term SeqPoints. We analyze two state-of-the-art SQNNs, DeepSpeech2 and Google's Neural Machine Translation (GNMT), and show that SeqPoints can represent their entire training runs accurately, resulting in geomean errors of only 0.11% and 0.53%, respectively, when projecting overall runtime and 0.13% and 1.50% when projecting speedups due to architectural changes. This high accuracy is achieved while reducing the time needed for profiling by 345x and 214x for the two networks compared to full training runs. As a result, SeqPoint can enable analysis of SQNN training runs in mere minutes instead of hours or days., Comment: To appear in IEEE International Symposium on Performance Analysis of Systems and Software (ISPASS 2020)

This work provides the first study to explore the interaction of update propagation with and without fine-grained synchronization (push vs. pull), emerging coherence protocols (GPU vs. DeNovo coherence), and software-centric consistency models (DRF0, DRF1, and DRFrlx) for graph workloads on emerging integrated GPU-CPU systems with native unified shared memory. We study 6 graph applications with 6 graph inputs for a total of 36 workloads running on 12 system (hardware+software) configurations reflecting the above design space of update propagation, coherence, and memory consistency. We make three key contributions. First, we show that there is no single best system configuration for all workloads, motivating systems with flexible coherence and consistency support. Second, we develop a model to accurately predict the best system configuration -- this model can be used by software designers to decide on push vs. pull and the consistency model and by flexible hardware to invoke the appropriate coherence and consistency configuration for the given workload. Third, we show that the design dimensions explored here are inter-dependent, reinforcing the need for software-hardware co-design in the above design dimensions. For example, software designers deciding on push vs. pull must consider the consistency model supported by hardware -- in some cases, push maybe better if hardware supports DRFrlx while pull may be better if hardware does not support DRFrlx.

Most deep neural networks deployed today are trained using GPUs via high-level frameworks such as TensorFlow and PyTorch. This paper describes changes we made to the GPGPU-Sim simulator to enable it to run PyTorch by running PTX kernels included in NVIDIA's cuDNN library. We use the resulting modified simulator, which has been made available publicly with this paper, to study some simple deep learning workloads. With our changes to GPGPU-Sim's functional simulation model, we find GPGPU-Sim performance model running a cuDNN enabled implementation of LeNet for MNIST reports results within 30% of real hardware. Using GPGPU-Sim's AerialVision performance analysis tool we observe that cuDNN API calls contain many varying phases and appear to include potentially inefficient microarchitecture behaviour such as DRAM partition bank camping, at least when executed on GPGPU-Sim's current performance model., Comment: Source code available at: https://github.com/gpgpu-sim/gpgpu-sim_distribution/tree/dev