GPUs power modern scientific and AI applications, but their limited memory capacity restricts scalability. Buying GPUs with larger HBM is prohibitively expensive and still bounded by market limits. Existing solutions either exploit application-specific knowledge through out-of-core techniques, which lack generality, or rely on system-level page faulting, which is transparent but inefficient. We propose FRUGAL, an application-agnostic framework and methodology that reduces GPU memory footprint while sustaining high performance. FRUGAL formulates memory management as an optimization over an application’s execution graph, encompassing prefetching, kernel execution, and offloading. Using static analysis and profiling, FRUGAL applies a two-phase scheduling and migration strategy, solving an otherwise intractable optimization efficiently. Evaluations on Tiled Cholesky Decomposition, Tiled LU Decomposition, Tiny-CUDA-NN, and QuEST show that FRUGAL significantly reduces maximum GPU memory usage by 80.21%, 80.20%, 64.75% and 60.86% with only a geometric mean of 28.31% slowdown. FRUGAL allows applications to exceed hardware-imposed limits, and maintains strong performance scalability beyond existing GPU memory constraints, without additional hardware cost.

Distributed Sparse Matrix-Matrix Multiplication (SpMM) is a fundamental operation in high-performance computing and deep learning applications. The major performance bottleneck in distributed SpMM lies in substantial communication overhead, which limits both performance and scalability. In this paper, we identify two key sources of communication inefficiency in distributed SpMM: redundant data transfer due to sparsity unawareness, and suboptimal utilization of hierarchical network topology. To address these, we propose (1) a fine-grained, sparsity-aware communication strategy that reduces communication overhead by exploiting the sparsity pattern of the sparse matrix, and (2) a hierarchical communication strategy that maps the sparsity-aware strategy onto two-tier GPU network architectures, minimizing redundant data movement across slower inter-node links. We implement these optimizations in a comprehensive distributed SpMM framework, SHIRO. Extensive evaluations on real-world datasets show that SHIRO demonstrates strong scalability up to 128 GPUs, achieving geometric mean speedups of 221.5 ×, 56.0 ×, 23.4 ×, and 8.8 × in SpMM over four state-of-the-art baselines (CAGNET, SPA, BCL, and CoLa, respectively) at this scale.

Graph Convolutional Networks (GCNs), particularly for largescale graphs, are crucial across numerous domains.However, training distributed full-batch GCNs on large-scale graphs suffers from inefficient memory access patterns and high communication overhead.To address these challenges, we introduce SuperGCN, an efficient and scalable distributed GCN

Graph Convolutional Networks (GCNs) are widely used in various domains. However, training distributed full-batch GCNs on large-scale graphs poses challenges due to inefficient memory access patterns and high communication overhead. This paper presents a general and efficient GCN training framework on CPU supercomputers. It comprises a general aggregation kernel designed to optimize irregular memory access and a quantization method with label propagation to reduce communication overhead. Experimental results show that our method achieves a speedup of up to 4.1× compared with the SoTA implementations.

Iterative memory-bound solvers commonly occur in HPC codes. Typical GPU implementations have a loop on the host side that invokes the GPU kernel as much as time/algorithm steps there are. The termination of each kernel implicitly acts the barrier required after advancing the solution every time step. We propose an execution model for running memory-bound iterative GPU kernels: PERsistent KernelS (PERKS). In this model, the time loop is moved inside persistent kernel, and device-wide barriers are used for synchronization. We then reduce the traffic to device memory by caching subset of the output in each time step in the unused registers and shared memory. PERKS can be generalized to any iterative solver: they largely independent of the solver's implementation. We explain the design principle of PERKS and demonstrate effectiveness of PERKS for a wide range of iterative 2D/3D stencil benchmarks (geomean speedup of 2.12x for 2D stencils and 1.24x for 3D stencils over state-of-art libraries), and a Krylov subspace conjugate gradient solver (geomean speedup of 4.86x in smaller SpMV datasets from SuiteSparse and 1.43x in larger SpMV datasets over a state-of-art library). All PERKS-based implementations available at: https://github.com/neozhang307/PERKS.

Iterative stencils are used widely across the spectrum of High Performance Computing (HPC) applications. Many efforts have been put into optimizing stencil GPU kernels, given the prevalence of GPU-accelerated supercomputers. To improve the data locality, temporal blocking is an optimization that combines a batch of time steps to process them together. Under the observation that GPUs are evolving to resemble CPUs in some aspects, we revisit temporal blocking optimizations for GPUs. We explore how temporal blocking schemes can be adapted to the new features in the recent Nvidia GPUs, including large scratchpad memory, hardware prefetching, and device-wide synchronization. We propose a novel temporal blocking method, EBISU, which champions low device occupancy to drive aggressive deep temporal blocking on large tiles that are executed tile-by-tile. We compare EBISU with state-of-the-art temporal blocking libraries: STENCILGEN and AN5D. We also compare with state-of-the-art stencil auto-tuning tools that are equipped with temporal blocking optimizations: ARTEMIS and DRSTENCIL. Over a wide range of stencil benchmarks, EBISU achieves speedups up to 2.53x and a geometric mean speedup of 1.49x over the best state-of-the-art performance in each stencil benchmark.

Over the last three decades, innovations in the memory subsystem were primarily targeted at overcoming the data movement bottleneck. In this paper, we focus on a specific market trend in memory technology: 3D-stacked memory and caches. We investigate the impact of extending the on-chip memory capabilities in future HPC-focused processors, particularly by 3D-stacked SRAM. First, we propose a method oblivious to the memory subsystem to gauge the upper-bound in performance improvements when data movement costs are eliminated. Then, using the gem5 simulator, we model two variants of a hypothetical LARge Cache processor (LARC), fabricated in 1.5 nm and enriched with high-capacity 3D-stacked cache. With a volume of experiments involving a broad set of proxy-applications and benchmarks, we aim to reveal how HPC CPU performance will evolve, and conclude an average boost of 9.56× for cache-sensitive HPC applications, on a per-chip basis. Additionally, we exhaustively document our methodological exploration to motivate HPC centers to drive their own technological agenda through enhanced co-design.

Matrix engines or units, in different forms and affinities, are becoming a reality in modern processors; CPUs and otherwise. The current and dominant algorithmic approach to Deep Learning merits the commercial investments in these units, and deduced from the No. 1 benchmark in supercomputing, namely High Performance Linpack, one would expect an awakened enthusiasm by the HPC community, too. Hence, our goal is to identify the practical added benefits for HPC and machine learning applications by having access to matrix engines. For this purpose, we perform an in-depth survey of software stacks, proxy applications and benchmarks, and historical batch job records. We provide a cost-benefit analysis of matrix engines, both asymptotically and in conjunction with state-of-the-art processors. While our empirical data will temper the enthusiasm, we also outline opportunities to “misuse” these dense matrix-multiplication engines if they come for free.

GPUs are playing an increasingly important role in general-purpose computing. Many algorithms require synchronizations at different levels of granularity in a single GPU. Additionally, the emergence of dense GPU nodes also calls for multi-GPU synchronization. Nvidia's latest CUDA provides a variety of synchronization methods. Until now, there is no full understanding of the characteristics of those synchronization methods. This work explores important undocumented features and provides an in-depth analysis of the performance considerations and pitfalls of the state-of-art synchronization methods for Nvidia GPUs. The provided analysis would be useful when making design choices for applications, libraries, and frameworks running on single and/or multi-GPU environments. We provide a case study of the commonly used reduction operator to illustrate how the knowledge gained in our analysis can be useful. We also describe our micro-benchmarks and measurement methods.

The dedicated memory of hardware accelerators can be insufficient to store all weights and/or intermediate states of large deep learning models. Although model parallelism is a viable approach to reduce the memory pressure issue, significant modification of the source code and considerations for algorithms are required. An alternative solution is to use out-of-core methods instead of, or in addition to, data parallelism. We propose a performance model based on the concurrency analysis of out-of-core training behavior, and derive a strategy that combines layer swapping and redundant recomputing. We achieve an average of 1. 52x speedup in six different models over the state-of-the-art out-of-core methods. We also introduce the first method to solve the challenging problem of out-of-core multi-node training by carefully pipelining gradient exchanges and performing the parameter updates on the host. Our data parallel out-of-core solution can outperform complex hybrid model parallelism in training large models, e.g. Megatron-LM and Turning-NLG.