More Like this

1. CheckFreq: Frequent, Fine-Grained DNN Checkpointing

   Mohan, Jayashree ; Phanishayee, Amar ; Chidambaram, Vijay ( February 2021 , Proceedings of the USENIX Conference on File and Storage Technologies (FAST 2021))

   Aguilera, Marcos ; Yadgar, Gala (Ed.)

   Training Deep Neural Networks (DNNs) is a resource-hungry and time-consuming task. During training, the model performs computation at the GPU to learn weights, repeatedly, over several epochs. The learned weights reside in GPU memory, and are occasionally checkpointed (written to persistent storage) for fault-tolerance. Traditionally, model parameters are checkpointed at epoch boundaries; for modern deep networks, an epoch runs for several hours. An interruption to the training job due to preemption, node failure, or process failure, therefore results in the loss of several hours worth of GPU work on recovery. We present CheckFreq, an automatic, fine-grained checkpointing framework that (1) algorithmically determines the checkpointing frequency at the granularity of iterations using systematic online profiling, (2) dynamically tunes checkpointing frequency at runtime to bound the checkpointing overhead using adaptive rate tuning, (3) maintains the training data invariant of using each item in the dataset exactly once per epoch by checkpointing data loader state using a light-weight resumable iterator, and (4) carefully pipelines checkpointing with computation to reduce the checkpoint cost by introducing two-phase checkpointing. Our experiments on a variety of models, storage backends, and GPU generations show that CheckFreq can reduce the recovery time from hours to seconds while bounding the runtime overhead within 3.5%. 

   more » « less
2. CRUM: Checkpoint-Restart Support for CUDA's Unified Memory

   https://doi.org/10.1109/CLUSTER.2018.00047  

   Garg, Rohan ; Mohan, Apoorve ; Sullivan, Michael ; Cooperman, Gene ( September 2018 , Proc. of IEEE Int. Conf. on Cluster Computing (Cluster'18))

   Unified Virtual Memory (UVM) was recently introduced with CUDA version 8 and the Pascal GPU. The older CUDA programming style is akin to older large-memory UNIX applications which used to directly load and unload memory segments. Newer CUDA programs have started taking advantage of UVM for the same reasons of superior programmability that UNIX applications long ago switched to assuming the presence of virtual memory. Therefore, checkpointing of UVM has become increasing important, especially as NVIDIA CUDA continues to gain wider popularity: 87 of the top 500 supercomputers in the latest listings use NVIDIA GPUs, with a current trend of ten additional NVIDIA-based supercomputers each year. A new scalable checkpointing mechanism, CRUM (Checkpoint-Restart for Unified Memory), is demonstrated for hybrid CUDA/MPI computations across multiple computer nodes. The support for UVM is particularly attractive for programs requiring more memory than resides on the GPU, since the alternative to UVM is for the application to directly copy memory between device and host. Furthermore, CRUM supports a fast, forked checkpointing, which mostly overlaps the CUDA computation with storage of the checkpoint image in stable storage. The runtime overhead of using CRUM is 6% on average, and the time for forked checkpointing is seen to be a factor of up to 40 times less than traditional, synchronous checkpointing. 

   more » « less
3. An autoencoder compression approach for accelerating large-scale inverse problems

   https://doi.org/10.1088/1361-6420/acfbe1  

   Wittmer, Jonathan ; Badger, Jacob ; Sundar, Hari ; Bui-Thanh, Tan ( October 2023 , Inverse Problems)

   N/A (Ed.)

   Partial differential equation (PDE)-constrained inverse problems are some of the most challenging and computationally demanding problems in computational science today. Fine meshes required to accurately compute the PDE solution introduce an enormous number of parameters and require large-scale computing resources such as more processors and more memory to solve such systems in a reasonable time. For inverse problems constrained by time-dependent PDEs, the adjoint method often employed to compute gradients and higher order derivatives efficiently requires solving a time-reversed, so-called adjoint PDE that depends on the forward PDE solution at each timestep. This necessitates the storage of a high-dimensional forward solution vector at every timestep. Such a procedure quickly exhausts the available memory resources. Several approaches that trade additional computation for reduced memory footprint have been proposed to mitigate the memory bottleneck, including checkpointing and compression strategies. In this work, we propose a close-to-ideal scalable compression approach using autoencoders to eliminate the need for checkpointing and substantial memory storage, thereby reducing the time-to-solution and memory requirements. We compare our approach with checkpointing and an off-the-shelf compression approach on an earth-scale ill-posed seismic inverse problem. The results verify the expected close-to-ideal speedup for the gradient and Hessian-vector product using the proposed autoencoder compression approach. To highlight the usefulness of the proposed approach, we combine the autoencoder compression with the data-informed active subspace (DIAS) prior showing how the DIAS method can be affordably extended to large-scale problems without the need for checkpointing and large memory.

    

   more » « less
4. Scalable Incremental Checkpointing using GPU-Accelerated De-Duplication

   https://doi.org/10.1145/3605573.3605639  

   Tan, Nigel ; Luettgau, Jakob ; Marquez, Jack ; Teranishi, Keita ; Morales, Nicolas ; Bhowmick, Sanjukta ; Cappello, Franck ; Taufer, Michela ; Nicolae, Bogdan ( August 2023 , ACM)

   Writing large amounts of data concurrently to stable storage is a typical I/O pattern of many HPC workflows. This pattern introduces high I/O overheads and results in increased storage space utilization especially for workflows that need to capture the evolution of data structures with high frequency as checkpoints. In this context, many applications, such as graph pattern matching, perform sparse updates to large data structures between checkpoints. For these applications, incremental checkpointing techniques that save only the differences from one checkpoint to another can dramatically reduce the checkpoint sizes, I/O bottlenecks, and storage space utilization. However, such techniques are not without challenges: it is non-trivial to transparently determine what data has changed since a previous checkpoint and assemble the differences in a compact fashion that does not result in excessive metadata. State-of-art data reduction techniques (e.g., compression and de-duplication) have significant limitations when applied to modern HPC applications that leverage GPUs: slow at detecting the differences, generate a large amount of metadata to keep track of the differences, and ignore crucial spatiotemporal checkpoint data redundancy. This paper addresses these challenges by proposing a Merkle tree-based incremental checkpointing method to exploit GPUs’ high memory bandwidth and massive parallelism. Experimental results at scale show a significant reduction of the I/O overhead and space utilization of checkpointing compared with state-of-the-art incremental checkpointing and compression techniques. 

   more » « less
5. Neural Network Compression for Noisy Storage Devices

   https://doi.org/10.1145/3588436  

   Isik, Berivan ; Choi, Kristy ; Zheng, Xin ; Weissman, Tsachy ; Ermon, Stefano ; Wong, H.-S. Philip ; Alaghi, Armin ( May 2023 , ACM Transactions on Embedded Computing Systems)

   Compression and efficient storage ofneural network (NN)parameters is critical for applications that run on resource-constrained devices. Despite the significant progress in NN model compression, there has been considerably less investigation in the actualphysicalstorage of NN parameters. Conventionally, model compression and physical storage are decoupled, as digital storage media witherror-correcting codes (ECCs)provide robust error-free storage. However, this decoupled approach is inefficient as it ignores the overparameterization present in most NNs and forces the memory device to allocate the same amount of resources to every bit of information regardless of its importance. In this work, we investigate analog memory devices as an alternative to digital media – one that naturally provides a way to add more protection for significant bits unlike its counterpart, but is noisy and may compromise the stored model’s performance if used naively. We develop a variety of robust coding strategies for NN weight storage on analog devices, and propose an approach to jointly optimize model compression and memory resource allocation. We then demonstrate the efficacy of our approach on models trained on MNIST, CIFAR-10, and ImageNet datasets for existing compression techniques. Compared to conventional error-free digital storage, our method reduces the memory footprint by up to one order of magnitude, without significantly compromising the stored model’s accuracy.

    

   more » « less