اشترك بالحزمة الذهبية واحصل على وصول غير محدود شمرا أكاديميا
تسجيل مستخدم جديدAccelerating SLIDE Deep Learning on Modern CPUs: Vectorization, Quantizations, Memory Optimizations, and More
164
0
0.0
(
0
)
اسأل ChatGPT حول البحث
ﻻ يوجد ملخص باللغة العربية
Deep learning implementations on CPUs (Central Processing Units) are gaining more traction. Enhanced AI capabilities on commodity x86 architectures are commercially appealing due to the reuse of existing hardware and virtualization ease. A notable work in this direction is the SLIDE system. SLIDE is a C++ implementation of a sparse hash table based back-propagation, which was shown to be significantly faster than GPUs in training hundreds of million parameter neural models. In this paper, we argue that SLIDEs current implementation is sub-optimal and does not exploit several opportunities available in modern CPUs. In particular, we show how SLIDEs computations allow for a unique possibility of vectorization via AVX (Advanced Vector Extensions)-512. Furthermore, we highlight opportunities for different kinds of memory optimization and quantizations. Combining all of them, we obtain up to 7x speedup in the computations on the same hardware. Our experiments are focused on large (hundreds of millions of parameters) recommendation and NLP models. Our work highlights several novel perspectives and opportunities for implementing randomized algorithms for deep learning on modern CPUs. We provide the code and benchmark scripts at https://github.com/RUSH-LAB/SLIDE
قيم البحث
اقرأ أيضاً
Deep Neural Networks (DNNs) are witnessing increased adoption in multiple domains owing to their high accuracy in solving real-world problems. However, this high accuracy has been achieved by building deeper networks, posing a fundamental challenge t
o the low latency inference desired by user-facing applications. Current low latency solutions trade-off on accuracy or fail to exploit the inherent temporal locality in prediction serving workloads. We observe that caching hidden layer outputs of the DNN can introduce a form of late-binding where inference requests only consume the amount of computation needed. This enables a mechanism for achieving low latencies, coupled with an ability to exploit temporal locality. However, traditional caching approaches incur high memory overheads and lookup latencies, leading us to design learned caches - caches that consist of simple ML models that are continuously updated. We present the design of GATI, an end-to-end prediction serving system that incorporates learned caches for low-latency DNN inference. Results show that GATI can reduce inference latency by up to 7.69X on realistic workloads.
Federated learning (FL) is becoming a popular paradigm for collaborative learning over distributed, private datasets owned by non-trusting entities. FL has seen successful deployment in production environments, and it has been adopted in services suc
h as virtual keyboards, auto-completion, item recommendation, and several IoT applications. However, FL comes with the challenge of performing training over largely heterogeneous datasets, devices, and networks that are out of the control of the centralized FL server. Motivated by this inherent setting, we make a first step towards characterizing the impact of device and behavioral heterogeneity on the trained model. We conduct an extensive empirical study spanning close to 1.5K unique configurations on five popular FL benchmarks. Our analysis shows that these sources of heterogeneity have a major impact on both model performance and fairness, thus sheds light on the importance of considering heterogeneity in FL system design.
Energy proportionality is the key design goal followed by architects of modern multicore CPUs. One of its implications is that optimization of an application for performance will also optimize it for energy. In this work, we show that energy proporti
onality does not hold true for multicore CPUs. This finding creates the opportunity for bi-objective optimization of applications for performance and energy. We propose and study the first application-level method for bi-objective optimization of multithreaded data-parallel applications for performance and energy. The method uses two decision variables, the number of identical multithreaded kernels (threadgroups) executing the application and the number of threads in each threadgroup, with the workload always partitioned equally between the threadgroups. We experimentally demonstrate the efficiency of the method using four highly optimized multithreaded data-parallel applications, 2D fast Fourier transform based on FFTW and Intel MKL, and dense matrix-matrix multiplication using OpenBLAS and Intel MKL. Four modern multicore CPUs are used in the experiments. The experiments show that optimization for performance alone results in the increase in dynamic energy consumption by up to 89% and optimization for dynamic energy alone degrades the performance by up to 49%. By solving the bi-objective optimization problem, the method determines up to 11 globally Pareto-optimal solutions. Finally, we propose a qualitative dynamic energy model employing performance monitoring counters as parameters, which we use to explain the discovered energy nonproportionality and the Pareto-optimal solutions determined by our method. The model shows that the energy nonproportionality in our case is due to the activity of the data translation lookaside buffer (dTLB), which is disproportionately energy expensive.
We describe the new field of mathematical analysis of deep learning. This field emerged around a list of research questions that were not answered within the classical framework of learning theory. These questions concern: the outstanding generalizat
ion power of overparametrized neural networks, the role of depth in deep architectures, the apparent absence of the curse of dimensionality, the surprisingly successful optimization performance despite the non-convexity of the problem, understanding what features are learned, why deep architectures perform exceptionally well in physical problems, and which fine aspects of an architecture affect the behavior of a learning task in which way. We present an overview of modern approaches that yield partial answers to these questions. For selected approaches, we describe the main ideas in more detail.
Federated Learning (FL) is a recent development in the field of machine learning that collaboratively trains models without the training data leaving client devices, to preserve data privacy. In realistic FL settings, the training set is distributed
over clients in a highly non-Independent and Identically Distributed (non-IID) fashion, which has been shown extensively to harm FL convergence speed and final model performance. To address this challenge, we propose a novel, generalised approach for incorporating adaptive optimisation techniques into FL with the Federated Global Biased Optimiser (FedGBO) algorithm. FedGBO accelerates FL by employing a set of global biased optimiser values during the client-training phase, which helps to reduce `client-drift from non-IID data, whilst also benefiting from adaptive optimisation. We show that the FedGBO update with a generic optimiser can be reformulated as centralised training using biased gradients and optimiser updates, and apply this theoretical framework to prove the convergence of FedGBO using momentum-Stochastic Gradient Descent (SGDm). We also conduct extensive experiments using 4 realistic FL benchmark datasets (CIFAR100, Sent140, FEMNIST, Shakespeare) and 3 popular adaptive optimisers (RMSProp, SGDm, Adam) to compare the performance of state-of-the-art adaptive-FL algorithms. The results demonstrate that FedGBO has highly competitive performance whilst achieving lower communication and computation costs, and provide practical insights into the trade-offs associated with the different adaptive-FL algorithms and optimisers for real-world FL deployments.
الأسئلة المقترحة
سجل دخول لتتمكن من نشر تعليقات
التعليقات
جاري جلب التعليقات
سجل دخول لتتمكن من متابعة معايير البحث التي قمت باختيارها
