Large-scale pretrained AI models have shown state-of-the-art accuracy in a series of important applications. As the size of pretrained AI models grows dramatically each year in an effort to achieve higher accuracy, training such models requires massive computing and memory capabilities, which accelerates the convergence of AI and HPC. However, there are still gaps in deploying AI applications on HPC systems, which need application and system co-design based on specific hardware features. To this end, this paper proposes BaGuaLu1, the first work targeting training brain scale models on an entire exascale supercomputer, the New Generation Sunway Supercomputer. By combining hardware-specific intra-node optimization and hybrid parallel strategies, BaGuaLu enables decent performance and scalability on unprecedentedly large models. The evaluation shows that BaGuaLu can train 14.5-trillion-parameter models with a performance of over 1 EFLOPS using mixed-precision and has the capability to train 174-trillion-parameter models, which rivals the number of synapses in a human brain.
Large language models (LLMs) have demonstrated powerful capabilities, requiring huge memory with their increasing sizes and sequence lengths, thus demanding larger parallel systems. The broadly adopted pipeline parallelism introduces even heavier and unbalanced memory consumption. Recomputation is a widely employed technique to mitigate the problem but introduces extra computation overhead. This paper proposes AdaPipe, which aims to find the optimized recomputation and pipeline stage partitioning strategy. AdaPipe employs adaptive recomputation to maximize memory utilization and reduce the computation cost of each pipeline stage. A flexible stage partitioning algorithm is also adopted to balance the computation between different stages. We evaluate AdaPipe by training two representative models, GPT-3 (175B) and Llama 2 (70B), achieving up to 1.32× and 1.22× speedup on clusters with NVIDIA GPUs and Ascend NPUs respectively.
Out-of-core systems rely on high-performance cache sub-systems to reduce the number of I/O operations. Although the page cache in modern operating systems enables transparent access to memory and storage devices, it suffers from efficiency and scalability issues on cache misses, forcing out-of-core systems to design and implement their own cache components, which is a non-trivial task. This study proposes TriCache, a cache mechanism that enables in-memory programs to efficiently process out-of-core datasets without requiring any code rewrite. It provides a virtual memory interface on top of the conventional block interface to simultaneously achieve user transparency and sufficient out-of-core performance. A multi-level block cache design is proposed to address the challenge of per-access address translations required by a memory interface. It can exploit spatial and temporal localities in memory or storage accesses to render storage-to-memory address translation and page-level concurrency control adequately efficient for the virtual memory interface. Our evaluation shows that in-memory systems operating on top of TriCache can outperform Linux OS page cache by more than one order of magnitude, and can deliver performance comparable to or even better than that of corresponding counterparts designed specifically for out-of-core scenarios.