Deep Neural Networks (DNNs) have demonstrated great success in many fields such as image recognition and text analysis. However, the ever-increasing sizes of both DNN models and training datasets make deep leaning extremely computation- and memory-intensive. Recently, photonic computing has emerged as a promising technology for accelerating DNNs. While the design of photonic accelerators for DNN inference and forward propagation of DNN training has been widely investigated, the architectural acceleration for equally important backpropagation of DNN training has not been well studied. In this paper, we propose a novel silicon photonic-based backpropagation accelerator for high performance DNN training. Specifically, a general-purpose photonic gradient descent unit named STADIA is designed to implement the multiplication, accumulation, and subtraction operations required for computing gradients using mature optical devices including Mach-Zehnder Interferometer (MZI) and Mircoring Resonator (MRR), which can significantly reduce the training latency and improve the energy efficiency of backpropagation. To demonstrate efficient parallel computing, we propose a STADIA-based backpropagation acceleration architecture and design a dataflow by using wavelength-division multiplexing (WDM). We analyze the precision of STADIA by quantifying the precision limitations imposed by losses and noises. Furthermore, we evaluate STADIA with different element sizes by analyzing the power, area and time delay for photonic accelerators based on DNN models such as AlexNet, VGG19 and ResNet. Simulation results show that the proposed architecture STADIA can achieve significant improvement by 9.7× in time efficiency and 147.2× in energy efficiency, compared with the most advanced optical-memristor based backpropagation accelerator.
{"title":"STADIA: Photonic Stochastic Gradient Descent for Neural Network Accelerators","authors":"Chengpeng Xia, Yawen Chen, Haibo Zhang, Jigang Wu","doi":"10.1145/3607920","DOIUrl":"https://doi.org/10.1145/3607920","url":null,"abstract":"Deep Neural Networks (DNNs) have demonstrated great success in many fields such as image recognition and text analysis. However, the ever-increasing sizes of both DNN models and training datasets make deep leaning extremely computation- and memory-intensive. Recently, photonic computing has emerged as a promising technology for accelerating DNNs. While the design of photonic accelerators for DNN inference and forward propagation of DNN training has been widely investigated, the architectural acceleration for equally important backpropagation of DNN training has not been well studied. In this paper, we propose a novel silicon photonic-based backpropagation accelerator for high performance DNN training. Specifically, a general-purpose photonic gradient descent unit named STADIA is designed to implement the multiplication, accumulation, and subtraction operations required for computing gradients using mature optical devices including Mach-Zehnder Interferometer (MZI) and Mircoring Resonator (MRR), which can significantly reduce the training latency and improve the energy efficiency of backpropagation. To demonstrate efficient parallel computing, we propose a STADIA-based backpropagation acceleration architecture and design a dataflow by using wavelength-division multiplexing (WDM). We analyze the precision of STADIA by quantifying the precision limitations imposed by losses and noises. Furthermore, we evaluate STADIA with different element sizes by analyzing the power, area and time delay for photonic accelerators based on DNN models such as AlexNet, VGG19 and ResNet. Simulation results show that the proposed architecture STADIA can achieve significant improvement by 9.7× in time efficiency and 147.2× in energy efficiency, compared with the most advanced optical-memristor based backpropagation accelerator.","PeriodicalId":50914,"journal":{"name":"ACM Transactions on Embedded Computing Systems","volume":"37 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2023-09-09","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"136108462","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":3,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Graph neural networks (GNNs) have emerged as a powerful approach for modelling and learning from graph-structured data. Multiple fields have since benefitted enormously from the capabilities of GNNs, such as recommendation systems, social network analysis, drug discovery, and robotics. However, accelerating and efficiently processing GNNs require a unique approach that goes beyond conventional artificial neural network accelerators, due to the substantial computational and memory requirements of GNNs. The slowdown of scaling in CMOS platforms also motivates a search for alternative implementation substrates. In this paper, we present GHOST , the first silicon-photonic hardware accelerator for GNNs. GHOST efficiently alleviates the costs associated with both vertex-centric and edge-centric operations. It implements separately the three main stages involved in running GNNs in the optical domain, allowing it to be used for the inference of various widely used GNN models and architectures, such as graph convolution networks and graph attention networks. Our simulation studies indicate that GHOST exhibits at least 10.2 × better throughput and 3.8 × better energy efficiency when compared to GPU, TPU, CPU and multiple state-of-the-art GNN hardware accelerators.
{"title":"GHOST: A Graph Neural Network Accelerator using Silicon Photonics","authors":"Salma Afifi, Febin Sunny, Amin Shafiee, Mahdi Nikdast, Sudeep Pasricha","doi":"10.1145/3609097","DOIUrl":"https://doi.org/10.1145/3609097","url":null,"abstract":"Graph neural networks (GNNs) have emerged as a powerful approach for modelling and learning from graph-structured data. Multiple fields have since benefitted enormously from the capabilities of GNNs, such as recommendation systems, social network analysis, drug discovery, and robotics. However, accelerating and efficiently processing GNNs require a unique approach that goes beyond conventional artificial neural network accelerators, due to the substantial computational and memory requirements of GNNs. The slowdown of scaling in CMOS platforms also motivates a search for alternative implementation substrates. In this paper, we present GHOST , the first silicon-photonic hardware accelerator for GNNs. GHOST efficiently alleviates the costs associated with both vertex-centric and edge-centric operations. It implements separately the three main stages involved in running GNNs in the optical domain, allowing it to be used for the inference of various widely used GNN models and architectures, such as graph convolution networks and graph attention networks. Our simulation studies indicate that GHOST exhibits at least 10.2 × better throughput and 3.8 × better energy efficiency when compared to GPU, TPU, CPU and multiple state-of-the-art GNN hardware accelerators.","PeriodicalId":50914,"journal":{"name":"ACM Transactions on Embedded Computing Systems","volume":"29 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2023-09-09","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"136108465","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":3,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Shaokai Lin, Yatin A. Manerkar, Marten Lohstroh, Elizabeth Polgreen, Sheng-Jung Yu, Chadlia Jerad, Edward A. Lee, Sanjit A. Seshia
Formal verification of cyber-physical systems (CPS) is challenging because it has to consider real-time and concurrency aspects that are often absent in ordinary software. Moreover, the software in CPS is often complex and low-level, making it hard to assure that a formal model of the system used for verification is a faithful representation of the actual implementation, which can undermine the value of a verification result. To address this problem, we propose a methodology for building verifiable CPS based on the principle that a formal model of the software can be derived automatically from its implementation. Our approach requires that the system implementation is specified in Lingua Franca (LF), a polyglot coordination language tailored for real-time, concurrent CPS, which we made amenable to the specification of safety properties via annotations in the code. The program structure and the deterministic semantics of LF enable automatic construction of formal axiomatic models directly from LF programs. The generated models are automatically checked using Bounded Model Checking (BMC) by the verification engine Uclid5 using the Z3 SMT solver. The proposed technique enables checking a well-defined fragment of Safety Metric Temporal Logic (Safety MTL) formulas. To ensure the completeness of BMC, we present a method to derive an upper bound on the completeness threshold of an axiomatic model based on the semantics of LF. We implement our approach in the LF V erifier and evaluate it using a benchmark suite with 22 programs sampled from real-life applications and benchmarks for Erlang, Lustre, actor-oriented languages, and RTOSes. The LF V erifier correctly checks 21 out of 22 programs automatically.
网络物理系统(CPS)的正式验证具有挑战性，因为它必须考虑在普通软件中经常缺失的实时和并发性方面。此外，CPS中的软件通常是复杂和低级的，这使得很难保证用于验证的系统的正式模型是实际实现的忠实表示，这可能会破坏验证结果的价值。为了解决这个问题，我们提出了一种构建可验证的CPS的方法，该方法基于软件的正式模型可以从其实现中自动导出的原则。我们的方法要求系统实现用Lingua Franca (LF)指定，这是一种为实时、并发CPS量身定制的多语言协调语言，我们通过代码中的注释使其符合安全属性的规范。LF的程序结构和确定性语义使得直接从LF程序自动构造形式化公理模型成为可能。生成的模型由验证引擎Uclid5使用Z3 SMT求解器自动使用BMC (Bounded Model Checking)进行检查。所提出的技术能够检查定义良好的安全度量时间逻辑(Safety MTL)公式片段。为了保证BMC的完备性，我们提出了一种基于LF语义的公理模型完备性阈值上界的推导方法。我们在LF V验证器中实现了我们的方法，并使用一个包含22个程序的基准测试套件来评估它，这些程序来自于现实生活中的应用程序和Erlang、Lustre、面向角色的语言和rtos的基准测试。LF V验证器自动正确检查22个程序中的21个。
{"title":"Towards Building Verifiable CPS using Lingua Franca","authors":"Shaokai Lin, Yatin A. Manerkar, Marten Lohstroh, Elizabeth Polgreen, Sheng-Jung Yu, Chadlia Jerad, Edward A. Lee, Sanjit A. Seshia","doi":"10.1145/3609134","DOIUrl":"https://doi.org/10.1145/3609134","url":null,"abstract":"Formal verification of cyber-physical systems (CPS) is challenging because it has to consider real-time and concurrency aspects that are often absent in ordinary software. Moreover, the software in CPS is often complex and low-level, making it hard to assure that a formal model of the system used for verification is a faithful representation of the actual implementation, which can undermine the value of a verification result. To address this problem, we propose a methodology for building verifiable CPS based on the principle that a formal model of the software can be derived automatically from its implementation. Our approach requires that the system implementation is specified in Lingua Franca (LF), a polyglot coordination language tailored for real-time, concurrent CPS, which we made amenable to the specification of safety properties via annotations in the code. The program structure and the deterministic semantics of LF enable automatic construction of formal axiomatic models directly from LF programs. The generated models are automatically checked using Bounded Model Checking (BMC) by the verification engine Uclid5 using the Z3 SMT solver. The proposed technique enables checking a well-defined fragment of Safety Metric Temporal Logic (Safety MTL) formulas. To ensure the completeness of BMC, we present a method to derive an upper bound on the completeness threshold of an axiomatic model based on the semantics of LF. We implement our approach in the LF V erifier and evaluate it using a benchmark suite with 22 programs sampled from real-life applications and benchmarks for Erlang, Lustre, actor-oriented languages, and RTOSes. The LF V erifier correctly checks 21 out of 22 programs automatically.","PeriodicalId":50914,"journal":{"name":"ACM Transactions on Embedded Computing Systems","volume":"77 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2023-09-09","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"136108727","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":3,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Physical Unclonable Functions (PUFs) have been widely considered an attractive security primitive. They use the deviations in the fabrication process to have unique responses from each device. Due to their nature, they serve as a DNA-like identity of the device. But PUFs have also been targeted for attacks. It has been proven that machine learning (ML) can be used to effectively model a PUF design and predict its behavior, leading to leakage of the internal secrets. To combat such attacks, several designs have been proposed to make it harder to model PUFs. One design direction is to use Non-Volatile Memory (NVM) as the building block of the PUF. NVM typically are multi-level cells, i.e, they have several internal states, which makes it harder to model them. However, the current state of the art of NVM-based PUFs is limited to ‘weak PUFs’, i.e., the number of outputs grows only linearly with the number of inputs, which limits the number of possible secret values that can be stored using the PUF. To overcome this limitation, in this work we design the Arbiter Non-Volatile PUF (ANV-PUF) that is exponential in the number of inputs and that is resilient against ML-based modeling. The concept is based on the famous delay-based Arbiter PUF (which is not resilient against modeling attacks) while using NVM as a building block instead of switches. Hence, we replace the switch delays (which are easy to model via ML) with the multi-level property of NVM (which is hard to model via ML). Consequently, our design has the exponential output characteristics of the Arbiter PUF and the resilience against attacks from the NVM-based PUFs. Our results show that the resilience to ML modeling, uniqueness, and uniformity are all in the ideal range of 50%. Thus, in contrast to the state-of-the-art, ANV-PUF is able to be resilient to attacks, while having an exponential number of outputs.
{"title":"ANV-PUF: Machine-Learning-Resilient NVM-Based Arbiter PUF","authors":"Hassan Nassar, Lars Bauer, Jörg Henkel","doi":"10.1145/3609388","DOIUrl":"https://doi.org/10.1145/3609388","url":null,"abstract":"Physical Unclonable Functions (PUFs) have been widely considered an attractive security primitive. They use the deviations in the fabrication process to have unique responses from each device. Due to their nature, they serve as a DNA-like identity of the device. But PUFs have also been targeted for attacks. It has been proven that machine learning (ML) can be used to effectively model a PUF design and predict its behavior, leading to leakage of the internal secrets. To combat such attacks, several designs have been proposed to make it harder to model PUFs. One design direction is to use Non-Volatile Memory (NVM) as the building block of the PUF. NVM typically are multi-level cells, i.e, they have several internal states, which makes it harder to model them. However, the current state of the art of NVM-based PUFs is limited to ‘weak PUFs’, i.e., the number of outputs grows only linearly with the number of inputs, which limits the number of possible secret values that can be stored using the PUF. To overcome this limitation, in this work we design the Arbiter Non-Volatile PUF (ANV-PUF) that is exponential in the number of inputs and that is resilient against ML-based modeling. The concept is based on the famous delay-based Arbiter PUF (which is not resilient against modeling attacks) while using NVM as a building block instead of switches. Hence, we replace the switch delays (which are easy to model via ML) with the multi-level property of NVM (which is hard to model via ML). Consequently, our design has the exponential output characteristics of the Arbiter PUF and the resilience against attacks from the NVM-based PUFs. Our results show that the resilience to ML modeling, uniqueness, and uniformity are all in the ideal range of 50%. Thus, in contrast to the state-of-the-art, ANV-PUF is able to be resilient to attacks, while having an exponential number of outputs.","PeriodicalId":50914,"journal":{"name":"ACM Transactions on Embedded Computing Systems","volume":"53 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2023-09-09","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"136192252","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":3,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
The continuous increase in data volume has led to the adoption of shingled-magnetic recording (SMR) as the primary technology for modern storage drives. This technology offers high storage density and low unit cost but introduces significant performance overheads due to the read-update-write operation and garbage collection (GC) process. To reduce these overheads, data deduplication has been identified as an effective solution as it reduces the amount of written data to an SMR-based storage device. However, deduplication can result in poor data locality, leading to decreased read performance. To tackle this problem, this study proposes a data locality-aware deduplication technology, LaDy, that considers both the overheads of writing duplicate data and the impact on data locality to determine whether the duplicate data should be written. LaDy integrates with DiskSim, an open-source project, and modifies it to simulate an SMR-based drive. The experimental results demonstrate that LaDy can significantly reduce the response time in the best-case scenario by 87.3% compared with CAFTL on the SMR drive. LaDy achieves this by selectively writing duplicate data, which preserves data locality, resulting in improved read performance. The proposed solution provides an effective and efficient method for mitigating the performance overheads associated with data deduplication in SMR-based storage devices.
{"title":"LaDy: Enabling <u>L</u> ocality- <u>a</u> ware <u>D</u> eduplication Technolog <u>y</u> on Shingled Magnetic Recording Drives","authors":"Jung-Hsiu Chang, Tzu-Yu Chang, Yi-Chao Shih, Tseng-Yi Chen","doi":"10.1145/3607921","DOIUrl":"https://doi.org/10.1145/3607921","url":null,"abstract":"The continuous increase in data volume has led to the adoption of shingled-magnetic recording (SMR) as the primary technology for modern storage drives. This technology offers high storage density and low unit cost but introduces significant performance overheads due to the read-update-write operation and garbage collection (GC) process. To reduce these overheads, data deduplication has been identified as an effective solution as it reduces the amount of written data to an SMR-based storage device. However, deduplication can result in poor data locality, leading to decreased read performance. To tackle this problem, this study proposes a data locality-aware deduplication technology, LaDy, that considers both the overheads of writing duplicate data and the impact on data locality to determine whether the duplicate data should be written. LaDy integrates with DiskSim, an open-source project, and modifies it to simulate an SMR-based drive. The experimental results demonstrate that LaDy can significantly reduce the response time in the best-case scenario by 87.3% compared with CAFTL on the SMR drive. LaDy achieves this by selectively writing duplicate data, which preserves data locality, resulting in improved read performance. The proposed solution provides an effective and efficient method for mitigating the performance overheads associated with data deduplication in SMR-based storage devices.","PeriodicalId":50914,"journal":{"name":"ACM Transactions on Embedded Computing Systems","volume":"1 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2023-09-09","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"136192262","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":3,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Control theory allows one to design controllers that are robust to external disturbances, model simplification, and modelling inaccuracy. Researchers have investigated whether the robustness carries on to the controller’s digital implementation, mostly looking at how the controller reacts to either communication or computational problems. Communication problems are typically modelled using random variables (i.e., estimating the probability that a fault will occur during a transmission), while computational problems are modelled using deterministic guarantees on the number of deadlines that the control task has to meet. These fault models allow the engineer to both design robust controllers and assess the controllers’ behaviour in the presence of isolated faults. Despite being very relevant for the real-world implementations of control system, the question of what happens when these faults occur simultaneously does not yet have a proper answer. In this paper, we answer this question in the stochastic setting, using the theory of Markov Jump Linear Systems to provide stability contracts with almost sure guarantees of convergence. For linear time-invariant Markov jump linear systems, mean square stability implies almost sure convergence – a property that is central to our investigation. Our research primarily emphasises the validation of this property for closed-loop systems that are subject to packet losses and computational overruns, potentially occurring simultaneously. We apply our method to two case studies from the recent literature and show their robustness to a comprehensive set of faults. We employ closed-loop system simulations to empirically derive performance metrics that elucidate the quality of the controller implementation, such as the system settling time and the integral absolute error.
{"title":"Stochastic Analysis of Control Systems Subject to Communication and Computation Faults","authors":"Nils Vreman, Martina Maggio","doi":"10.1145/3609123","DOIUrl":"https://doi.org/10.1145/3609123","url":null,"abstract":"Control theory allows one to design controllers that are robust to external disturbances, model simplification, and modelling inaccuracy. Researchers have investigated whether the robustness carries on to the controller’s digital implementation, mostly looking at how the controller reacts to either communication or computational problems. Communication problems are typically modelled using random variables (i.e., estimating the probability that a fault will occur during a transmission), while computational problems are modelled using deterministic guarantees on the number of deadlines that the control task has to meet. These fault models allow the engineer to both design robust controllers and assess the controllers’ behaviour in the presence of isolated faults. Despite being very relevant for the real-world implementations of control system, the question of what happens when these faults occur simultaneously does not yet have a proper answer. In this paper, we answer this question in the stochastic setting, using the theory of Markov Jump Linear Systems to provide stability contracts with almost sure guarantees of convergence. For linear time-invariant Markov jump linear systems, mean square stability implies almost sure convergence – a property that is central to our investigation. Our research primarily emphasises the validation of this property for closed-loop systems that are subject to packet losses and computational overruns, potentially occurring simultaneously. We apply our method to two case studies from the recent literature and show their robustness to a comprehensive set of faults. We employ closed-loop system simulations to empirically derive performance metrics that elucidate the quality of the controller implementation, such as the system settling time and the integral absolute error.","PeriodicalId":50914,"journal":{"name":"ACM Transactions on Embedded Computing Systems","volume":"27 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2023-09-09","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"136192421","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":3,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Intermittent deep neural network (DNN) inference is a promising technique to enable intelligent applications on tiny devices powered by ambient energy sources. Nonetheless, intermittent execution presents inherent challenges, primarily involving accumulating progress across power cycles and having to refetch volatile data lost due to power loss in each power cycle. Existing approaches typically optimize the inference configuration to maximize data reuse. However, we observe that such a fixed configuration may be significantly inefficient due to the fluctuating balance point between data reuse and data refetch caused by the dynamic nature of ambient energy. This work proposes DynBal , an approach to dynamically reconfigure the inference engine at runtime. DynBal is realized as a middleware plugin that improves inference performance by exploring the interplay between data reuse and data refetch to maintain their balance with respect to the changing level of intermittency. An indirect metric is developed to easily evaluate an inference configuration considering the variability in intermittency, and a lightweight reconfiguration algorithm is employed to efficiently optimize the configuration at runtime. We evaluate the improvement brought by integrating DynBal into a recent intermittent inference approach that uses a fixed configuration. Evaluations were conducted on a Texas Instruments device with various network models and under varied intermittent power strengths. Our experimental results demonstrate that DynBal can speed up intermittent inference by 3.26 times, achieving a greater improvement for a large network under high intermittency and a large gap between memory and computation performance.
{"title":"Keep in Balance: Runtime-reconfigurable Intermittent Deep Inference","authors":"Chih-Hsuan Yen, Hashan Roshantha Mendis, Tei-Wei Kuo, Pi-Cheng Hsiu","doi":"10.1145/3607918","DOIUrl":"https://doi.org/10.1145/3607918","url":null,"abstract":"Intermittent deep neural network (DNN) inference is a promising technique to enable intelligent applications on tiny devices powered by ambient energy sources. Nonetheless, intermittent execution presents inherent challenges, primarily involving accumulating progress across power cycles and having to refetch volatile data lost due to power loss in each power cycle. Existing approaches typically optimize the inference configuration to maximize data reuse. However, we observe that such a fixed configuration may be significantly inefficient due to the fluctuating balance point between data reuse and data refetch caused by the dynamic nature of ambient energy. This work proposes DynBal , an approach to dynamically reconfigure the inference engine at runtime. DynBal is realized as a middleware plugin that improves inference performance by exploring the interplay between data reuse and data refetch to maintain their balance with respect to the changing level of intermittency. An indirect metric is developed to easily evaluate an inference configuration considering the variability in intermittency, and a lightweight reconfiguration algorithm is employed to efficiently optimize the configuration at runtime. We evaluate the improvement brought by integrating DynBal into a recent intermittent inference approach that uses a fixed configuration. Evaluations were conducted on a Texas Instruments device with various network models and under varied intermittent power strengths. Our experimental results demonstrate that DynBal can speed up intermittent inference by 3.26 times, achieving a greater improvement for a large network under high intermittency and a large gap between memory and computation performance.","PeriodicalId":50914,"journal":{"name":"ACM Transactions on Embedded Computing Systems","volume":"28 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2023-09-09","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"136192113","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":3,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Neuromorphic computing is an emerging field with the potential to offer performance and energy-efficiency gains over traditional machine learning approaches. Most neuromorphic hardware, however, has been designed with limited concerns to the problem of integrating it with other components in a heterogeneous System-on-Chip (SoC). Building on a state-of-the-art reconfigurable neuromorphic architecture, we present the design of a neuromorphic hardware accelerator equipped with a programmable interface that simplifies both the integration into an SoC and communication with the processor present on the SoC. To optimize the allocation of on-chip resources, we develop an optimizer to restructure existing neuromorphic models for a given hardware architecture, and perform design-space exploration to find highly efficient implementations. We conduct experiments with various FPGA-based prototypes of many-accelerator SoCs, where Linux-based applications running on a RISC-V processor invoke Pareto-optimal implementations of our accelerator alongside third-party accelerators. These experiments demonstrate that our neuromorphic hardware, which is up to 89× faster and 170× more energy efficient after applying our optimizer, can be used in synergy with other accelerators for different application purposes.
{"title":"SpikeHard: Efficiency-Driven Neuromorphic Hardware for Heterogeneous Systems-on-Chip","authors":"Judicael Clair, Guy Eichler, Luca P. Carloni","doi":"10.1145/3609101","DOIUrl":"https://doi.org/10.1145/3609101","url":null,"abstract":"Neuromorphic computing is an emerging field with the potential to offer performance and energy-efficiency gains over traditional machine learning approaches. Most neuromorphic hardware, however, has been designed with limited concerns to the problem of integrating it with other components in a heterogeneous System-on-Chip (SoC). Building on a state-of-the-art reconfigurable neuromorphic architecture, we present the design of a neuromorphic hardware accelerator equipped with a programmable interface that simplifies both the integration into an SoC and communication with the processor present on the SoC. To optimize the allocation of on-chip resources, we develop an optimizer to restructure existing neuromorphic models for a given hardware architecture, and perform design-space exploration to find highly efficient implementations. We conduct experiments with various FPGA-based prototypes of many-accelerator SoCs, where Linux-based applications running on a RISC-V processor invoke Pareto-optimal implementations of our accelerator alongside third-party accelerators. These experiments demonstrate that our neuromorphic hardware, which is up to 89× faster and 170× more energy efficient after applying our optimizer, can be used in synergy with other accelerators for different application purposes.","PeriodicalId":50914,"journal":{"name":"ACM Transactions on Embedded Computing Systems","volume":"26 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2023-09-09","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"136107346","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":3,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Yongchun Zheng, Changlong Li, Yi Xiong, Weihong Liu, Cheng Ji, Zongwei Zhu, Lichen Yu
To ensure the user experience of mobile systems, the foreground application can be differentiated to minimize the impact of background applications. However, this article observes that system services in the kernel and framework layer, instead of background applications, are now the major resource competitors. Specifically, these service tasks tend to be quiet when people rarely interact with the foreground application and active when interactions become frequent, and this high overlap of busy times leads to contention for resources. This article proposes iAware, an interaction-aware task scheduling framework in mobile systems. The key insight is to make use of the previously ignored idle period and schedule service tasks to run at that period. iAware quantify the interaction characteristic based on the screen touch event, and successfully stagger the periods of frequent user interactions. With iAware, service tasks tend to run when few interactions occur, for example, when the device’s screen is turned off, instead of when the user is frequently interacting with it. iAware is implemented on real smartphones. Experimental results show that the user experience is significantly improved with iAware. Compared to the state-of-the-art, the application launching speed and frame rate are enhanced by 38.89% and 7.97% separately, with no more than 1% additional battery consumption.
{"title":"iAware: Interaction Aware Task Scheduling for Reducing Resource Contention in Mobile Systems","authors":"Yongchun Zheng, Changlong Li, Yi Xiong, Weihong Liu, Cheng Ji, Zongwei Zhu, Lichen Yu","doi":"10.1145/3609391","DOIUrl":"https://doi.org/10.1145/3609391","url":null,"abstract":"To ensure the user experience of mobile systems, the foreground application can be differentiated to minimize the impact of background applications. However, this article observes that system services in the kernel and framework layer, instead of background applications, are now the major resource competitors. Specifically, these service tasks tend to be quiet when people rarely interact with the foreground application and active when interactions become frequent, and this high overlap of busy times leads to contention for resources. This article proposes iAware, an interaction-aware task scheduling framework in mobile systems. The key insight is to make use of the previously ignored idle period and schedule service tasks to run at that period. iAware quantify the interaction characteristic based on the screen touch event, and successfully stagger the periods of frequent user interactions. With iAware, service tasks tend to run when few interactions occur, for example, when the device’s screen is turned off, instead of when the user is frequently interacting with it. iAware is implemented on real smartphones. Experimental results show that the user experience is significantly improved with iAware. Compared to the state-of-the-art, the application launching speed and frame rate are enhanced by 38.89% and 7.97% separately, with no more than 1% additional battery consumption.","PeriodicalId":50914,"journal":{"name":"ACM Transactions on Embedded Computing Systems","volume":"25 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2023-09-09","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"136107493","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":3,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Artem Klashtorny, Zhuanhao Wu, Anirudh Mohan Kaushik, Hiren Patel
We present a predictable wavefront splitting (PWS) technique for graphics processing units (GPUs). PWS improves the performance of GPU applications by reducing the impact of branch divergence while ensuring that worst-case execution time (WCET) estimates can be computed. This makes PWS an appropriate technique to use in safety-critical applications, such as autonomous driving systems, avionics, and space, that require strict temporal guarantees. In developing PWS on an AMD-based GPU, we propose microarchitectural enhancements to the GPU, and a compiler pass that eliminates branch serializations to reduce the WCET of a wavefront. Our analysis of PWS exhibits a performance improvement of 11% over existing architectures with a lower WCET than prior works in wavefront splitting.
{"title":"Predictable GPU Wavefront Splitting for Safety-Critical Systems","authors":"Artem Klashtorny, Zhuanhao Wu, Anirudh Mohan Kaushik, Hiren Patel","doi":"10.1145/3609102","DOIUrl":"https://doi.org/10.1145/3609102","url":null,"abstract":"We present a predictable wavefront splitting (PWS) technique for graphics processing units (GPUs). PWS improves the performance of GPU applications by reducing the impact of branch divergence while ensuring that worst-case execution time (WCET) estimates can be computed. This makes PWS an appropriate technique to use in safety-critical applications, such as autonomous driving systems, avionics, and space, that require strict temporal guarantees. In developing PWS on an AMD-based GPU, we propose microarchitectural enhancements to the GPU, and a compiler pass that eliminates branch serializations to reduce the WCET of a wavefront. Our analysis of PWS exhibits a performance improvement of 11% over existing architectures with a lower WCET than prior works in wavefront splitting.","PeriodicalId":50914,"journal":{"name":"ACM Transactions on Embedded Computing Systems","volume":"87 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2023-09-09","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"136107603","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":3,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
京公网安备 11010802042870号
京ICP备2023020795号-1
扫码关注我们
求助内容：
应助结果提醒方式：