An official website of the United States government
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
CheckFreq: Frequent, Fine-Grained DNN Checkpointing
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
- Award ID(s):
- 1751277
- PAR ID:
- 10286595
- Editor(s):
- Aguilera, Marcos; Yadgar, Gala
- Date Published:
- Journal Name:
- Proceedings of the USENIX Conference on File and Storage Technologies (FAST 2021)
- Format(s):
- Medium: X
- Sponsoring Org:
- National Science Foundation
More Like this
-
-
The likelihood of encountering in-training failures rises substantially with larger Deep Learning (DL) training workloads, leading to lost work and resource wastage. Such failures are typically offset by checkpointing, which comes at the cost of storage and network bandwidth overhead. State-of-the-art approaches involve lossy model compression mechanisms, which induce a tradeoff between the resulting model quality and compression ratio. We make a key enabling observation that the sensitivity of model weights to compression varies during training, and different weights benefit from different quantization levels, ranging from retaining full precision to pruning. We propose (1) a non-uniform quantization scheme that leverages this variation, (2) an efficient search mechanism that dynamically finds the best quantization configurations, and (3) a quantization-aware delta compression mechanism that rearranges weights to minimize checkpoint differences and thereby improving compression. We instantiate these contributions in Inshrinkerator, an in-training checkpoint compression system for DL workloads. Our experiments show that Inshrinkerator consistently achieves a better tradeoff between accuracy and compression ratio compared to prior works, enabling a compression ratio up to 39x and withstanding up to 10 restores with negligible accuracy impact in fault-tolerant training. Inshrinkerator achieves at least an order of magnitude reduction in checkpoint size for failure recovery and transfer learning without any loss of accuracy.more » « less
-
We consider the problem of orchestrating the execution of workflow applications structured as Directed Acyclic Graphs (DAGs) on parallel computing platforms that are subject to fail-stop failures. The objective is to minimize expected overall execution time, or makespan. A solution to this problem consists of a schedule of the workflow tasks on the available processors and of a decision of which application data to checkpoint to stable storage, so as to mitigate the impact of processor failures. For general DAGs this problem is hopelessly intractable. In fact, given a solution, computing its expected makespan is still a difficult problem. To address this challenge, we consider a restricted class of graphs, Minimal Series-Parallel Graphs (M-SPGS). It turns out that many real-world workflow applications are naturally structured as M-SPGS. For this class of graphs, we propose a recursive list-scheduling algorithm that exploits the M-SPG structure to assign sub-graphs to individual processors, and uses dynamic programming to decide which tasks in these sub-gaphs should be checkpointed. Furthermore, it is possible to efficiently compute the expected makespan for the solution produced by this algorithm, using a first-order approximation of task weights and existing evaluation algorithms for 2-state probabilistic DAGs. We assess the performance of our algorithm for production workflow configurations, comparing it to (i) an approach in which all application data is checkpointed, which corresponds to the standard way in which most production workflows are executed today; and (ii) an approach in which no application data is checkpointed. Our results demonstrate that our algorithm strikes a good compromise between these two approaches, leading to lower checkpointing overhead than the former and to better resilience to failure than the latter.more » « less
-
In the age of big data, deep learning has emerged as a powerful tool to extract insight and exploit its value, both in industry and scientific applications. One common pattern emerging in such applications is frequent checkpointing of the state of the learning model during training, needed in a variety of scenarios: analysis of intermediate states to explain features and correlations with training data, exploration strategies involving alternative models that share a common ancestor, knowledge transfer, resilience, etc. However, with increasing size of the learning models and popularity of distributed data-parallel training approaches, simple checkpointing techniques used so far face several limitations: low serialization performance, blocking I/O, stragglers due to the fact that only a single process is involved in checkpointing. This paper proposes a checkpointing technique specifically designed to address the aforementioned limitations, introducing efficient asynchronous techniques to hide the overhead of serialization and I/O, and distribute the load over all participating processes. Experiments with two deep learning applications (CANDLE and ResNet) on a pre-Exascale HPC platform (Theta) shows significant improvement over state-of-art, both in terms of checkpointing duration and runtime overhead.more » « less
-
Optimizing performance for Distributed Deep Neural Network (DDNN) training has recently become increasingly compelling, as the DNN model gets complex and the training dataset grows large. While existing works on communication scheduling mostly focus on overlapping the computation and communication to improve DDNN training performance, the GPU and network resources are still under-utilized in DDNN training clusters. To tackle this issue, in this paper, we design and implement a predictable communication scheduling strategy named Prophet to schedule the gradient transfer in an adequate order, with the aim of maximizing the GPU and network resource utilization. Leveraging our observed stepwise pattern of gradient transfer start time, Prophet first uses the monitored network bandwidth and the profiled time interval among gradients to predict the appropriate number of gradients that can be grouped into blocks. Then, these gradient blocks can be transferred one by one to guarantee high utilization of GPU and network resources while ensuring the priority of gradient transfer (i.e., low-priority gradients cannot preempt high-priority gradients in the network transfer). Prophet can make the forward propagation start as early as possible so as to greedily reduce the waiting (idle) time of GPU resources during the DDNN training process. Prototype experiments with representative DNN models trained on Amazon EC2 demonstrate that Prophet can improve the DDNN training performance by up to 40% compared with the state-of-theart priority-based communication scheduling strategies, yet with negligible runtime performance overhead.more » « less
- Free Publicly Accessible Full Text
-
Accepted Manuscript1.0
- Conference Paper:
- The DOI is not currently available.
