N. Mhaisen, Abhishek Sinha, G. Paschos, Georgios Iosifidis
We take a systematic look at the problem of storing whole files in a cache with limited capacity in the context of optimistic learning, where the caching policy has access to a prediction oracle (provided by, e.g., a Neural Network). The successive file requests are assumed to be generated by an adversary, and no assumption is made on the accuracy of the oracle. In this setting, we provide a universal lower bound for prediction-assisted online caching and proceed to design a suite of policies with a range of performance-complexity trade-offs. All proposed policies offer sublinear regret bounds commensurate with the accuracy of the oracle. Our results substantially improve upon all recently-proposed online caching policies, which, being unable to exploit the oracle predictions, offer only O(√T) regret. In this pursuit, we design, to the best of our knowledge, the first comprehensive optimistic Follow-the-Perturbed leader policy, which generalizes beyond the caching problem. We also study the problem of caching files with different sizes and the bipartite network caching problem. Finally, we evaluate the efficacy of the proposed policies through extensive numerical experiments using real-world traces.
{"title":"Optimistic No-regret Algorithms for Discrete Caching","authors":"N. Mhaisen, Abhishek Sinha, G. Paschos, Georgios Iosifidis","doi":"10.1145/3570608","DOIUrl":"https://doi.org/10.1145/3570608","url":null,"abstract":"We take a systematic look at the problem of storing whole files in a cache with limited capacity in the context of optimistic learning, where the caching policy has access to a prediction oracle (provided by, e.g., a Neural Network). The successive file requests are assumed to be generated by an adversary, and no assumption is made on the accuracy of the oracle. In this setting, we provide a universal lower bound for prediction-assisted online caching and proceed to design a suite of policies with a range of performance-complexity trade-offs. All proposed policies offer sublinear regret bounds commensurate with the accuracy of the oracle. Our results substantially improve upon all recently-proposed online caching policies, which, being unable to exploit the oracle predictions, offer only O(√T) regret. In this pursuit, we design, to the best of our knowledge, the first comprehensive optimistic Follow-the-Perturbed leader policy, which generalizes beyond the caching problem. We also study the problem of caching files with different sizes and the bipartite network caching problem. Finally, we evaluate the efficacy of the proposed policies through extensive numerical experiments using real-world traces.","PeriodicalId":426760,"journal":{"name":"Proceedings of the ACM on Measurement and Analysis of Computing Systems","volume":"18 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2022-08-15","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"125227898","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
We study the fairness of dynamic resource allocation problem under the α-fairness criterion. We recognize two different fairness objectives that naturally arise in this problem: the well-understood slot-fairness objective that aims to ensure fairness at every timeslot, and the less explored horizon-fairness objective that aims to ensure fairness across utilities accumulated over a time horizon. We argue that horizon-fairness comes at a lower price in terms of social welfare. We study horizon-fairness with the regret as a performance metric and show that vanishing regret cannot be achieved in presence of an unrestricted adversary. We propose restrictions on the adversary's capabilities corresponding to realistic scenarios and an online policy that indeed guarantees vanishing regret under these restrictions. We demonstrate the applicability of the proposed fairness framework to a representative resource management problem considering a virtualized caching system where different caches cooperate to serve content requests.
{"title":"Enabling Long-term Fairness in Dynamic Resource Allocation","authors":"T. Si Salem, G. Iosifidis, G. Neglia","doi":"10.1145/3570606","DOIUrl":"https://doi.org/10.1145/3570606","url":null,"abstract":"We study the fairness of dynamic resource allocation problem under the α-fairness criterion. We recognize two different fairness objectives that naturally arise in this problem: the well-understood slot-fairness objective that aims to ensure fairness at every timeslot, and the less explored horizon-fairness objective that aims to ensure fairness across utilities accumulated over a time horizon. We argue that horizon-fairness comes at a lower price in terms of social welfare. We study horizon-fairness with the regret as a performance metric and show that vanishing regret cannot be achieved in presence of an unrestricted adversary. We propose restrictions on the adversary's capabilities corresponding to realistic scenarios and an online policy that indeed guarantees vanishing regret under these restrictions. We demonstrate the applicability of the proposed fairness framework to a representative resource management problem considering a virtualized caching system where different caches cooperate to serve content requests.","PeriodicalId":426760,"journal":{"name":"Proceedings of the ACM on Measurement and Analysis of Computing Systems","volume":"150 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2022-08-11","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"115164988","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Recently it was shown that, contrary to expectations, the First-Come-First-Served (FCFS) scheduling algorithm can be stochastically improved upon by a scheduling algorithm called Nudge for light-tailed job size distributions. Nudge partitions jobs into 4 types based on their size, say small, medium, large and huge jobs. Nudge operates identical to FCFS, except that whenever a small job arrives that finds a large job waiting at the back of the queue, Nudge swaps the small job with the large one unless the large job was already involved in an earlier swap. In this paper, we show that FCFS can be stochastically improved upon under far weaker conditions. We consider a system with 2 job types and limited swapping between type-1 and type-2 jobs, but where a type-1 job is not necessarily smaller than a type-2 job. More specifically, we introduce and study the Nudge-K scheduling algorithm which allows type-1 jobs to be swapped with up to K type-2 jobs waiting at the back of the queue, while type-2 jobs can be involved in at most one swap. We present an explicit expression for the response time distribution under Nudge-K when both job types follow a phase-type distribution. Regarding the asymptotic tail improvement ratio (ATIR), we derive a simple expression for the ATIR, as well as for the K that maximizes the ATIR. We show that the ATIR is positive and the optimal K tends to infinity in heavy traffic as long as the type-2 jobs are on average longer than the type-1 jobs.
{"title":"On the Stochastic and Asymptotic Improvement of First-Come First-Served and Nudge Scheduling","authors":"B. Van Houdt","doi":"10.1145/3570610","DOIUrl":"https://doi.org/10.1145/3570610","url":null,"abstract":"Recently it was shown that, contrary to expectations, the First-Come-First-Served (FCFS) scheduling algorithm can be stochastically improved upon by a scheduling algorithm called Nudge for light-tailed job size distributions. Nudge partitions jobs into 4 types based on their size, say small, medium, large and huge jobs. Nudge operates identical to FCFS, except that whenever a small job arrives that finds a large job waiting at the back of the queue, Nudge swaps the small job with the large one unless the large job was already involved in an earlier swap. In this paper, we show that FCFS can be stochastically improved upon under far weaker conditions. We consider a system with 2 job types and limited swapping between type-1 and type-2 jobs, but where a type-1 job is not necessarily smaller than a type-2 job. More specifically, we introduce and study the Nudge-K scheduling algorithm which allows type-1 jobs to be swapped with up to K type-2 jobs waiting at the back of the queue, while type-2 jobs can be involved in at most one swap. We present an explicit expression for the response time distribution under Nudge-K when both job types follow a phase-type distribution. Regarding the asymptotic tail improvement ratio (ATIR), we derive a simple expression for the ATIR, as well as for the K that maximizes the ATIR. We show that the ATIR is positive and the optimal K tends to infinity in heavy traffic as long as the type-2 jobs are on average longer than the type-1 jobs.","PeriodicalId":426760,"journal":{"name":"Proceedings of the ACM on Measurement and Analysis of Computing Systems","volume":"69 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2022-06-21","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"121022900","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
The practicality of reinforcement learning algorithms has been limited due to poor scaling with respect to the problem size, as the sample complexity of learning an ε-optimal policy is Ω(|S||A|H/ ε2) over worst case instances of an MDP with state space S, action space A, and horizon H. We consider a class of MDPs for which the associated optimal Q* function is low rank, where the latent features are unknown. While one would hope to achieve linear sample complexity in |S| and |A| due to the low rank structure, we show that without imposing further assumptions beyond low rank of Q*, if one is constrained to estimate the Q function using only observations from a subset of entries, there is a worst case instance in which one must incur a sample complexity exponential in the horizon H to learn a near optimal policy. We subsequently show that under stronger low rank structural assumptions, given access to a generative model, Low Rank Monte Carlo Policy Iteration (LR-MCPI) and Low Rank Empirical Value Iteration (LR-EVI) achieve the desired sample complexity of Õ((|S|+|A|)poly (d,H)/ε2) for a rank d setting, which is minimax optimal with respect to the scaling of |S|, |A|, and ε. In contrast to literature on linear and low-rank MDPs, we do not require a known feature mapping, our algorithm is computationally simple, and our results hold for long time horizons. Our results provide insights on the minimal low-rank structural assumptions required on the MDP with respect to the transition kernel versus the optimal action-value function.
强化学习算法的实用性受到问题规模的限制，因为在具有状态空间S、动作空间A和视界H的MDP的最坏情况下，学习ε-最优策略的样本复杂度为Ω(|S| A|H/ ε2)。我们考虑一类MDP，其相关的最优Q*函数是低秩的，其中潜在特征是未知的。虽然由于低秩结构，人们希望在|S|和|A|中实现线性样本复杂度，但我们表明，在不施加超出低秩Q*的进一步假设的情况下，如果一个人被约束仅使用来自一个子集的观察值来估计Q函数，则存在最坏的情况，其中必须在视界H中产生样本复杂度指数来学习接近最优策略。我们随后证明，在更强的低秩结构假设下，给定生成模型，低秩蒙特卡罗策略迭代(LR-MCPI)和低秩经验值迭代(LR-EVI)对于秩d设置实现了所需的样本复杂度Õ((|S|+| a |)poly (d,H)/ε2)，该复杂度相对于|S|， | a |和ε的尺度是最小最大最优的。与线性和低秩mdp的文献相比，我们不需要已知的特征映射，我们的算法计算简单，我们的结果适用于很长的时间范围。我们的结果提供了关于相对于转移核和最优动作值函数的MDP所需的最小低秩结构假设的见解。
{"title":"Overcoming the Long Horizon Barrier for Sample-Efficient Reinforcement Learning with Latent Low-Rank Structure","authors":"Tyler Sam, Yudong Chen, C. Yu","doi":"10.1145/3589973","DOIUrl":"https://doi.org/10.1145/3589973","url":null,"abstract":"The practicality of reinforcement learning algorithms has been limited due to poor scaling with respect to the problem size, as the sample complexity of learning an ε-optimal policy is Ω(|S||A|H/ ε2) over worst case instances of an MDP with state space S, action space A, and horizon H. We consider a class of MDPs for which the associated optimal Q* function is low rank, where the latent features are unknown. While one would hope to achieve linear sample complexity in |S| and |A| due to the low rank structure, we show that without imposing further assumptions beyond low rank of Q*, if one is constrained to estimate the Q function using only observations from a subset of entries, there is a worst case instance in which one must incur a sample complexity exponential in the horizon H to learn a near optimal policy. We subsequently show that under stronger low rank structural assumptions, given access to a generative model, Low Rank Monte Carlo Policy Iteration (LR-MCPI) and Low Rank Empirical Value Iteration (LR-EVI) achieve the desired sample complexity of Õ((|S|+|A|)poly (d,H)/ε2) for a rank d setting, which is minimax optimal with respect to the scaling of |S|, |A|, and ε. In contrast to literature on linear and low-rank MDPs, we do not require a known feature mapping, our algorithm is computationally simple, and our results hold for long time horizons. Our results provide insights on the minimal low-rank structural assumptions required on the MDP with respect to the transition kernel versus the optimal action-value function.","PeriodicalId":426760,"journal":{"name":"Proceedings of the ACM on Measurement and Analysis of Computing Systems","volume":"64 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2022-06-07","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"133071796","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Residential Internet speeds have been rapidly increasing, reaching averages of ~100 Mbps in most developed countries. Several studies have shown that users have way more bandwidth than they need, only using about 20-30% on a regular day. Several systems exploit this trend by enabling users to monetize their spare bandwidth, e.g., by sharing their WiFi connection or by participating in distributed proxy or VPN (dVPN) services. Despite the proliferation of such systems, little is known on how such marketplaces operate, what are the key factors that determine the price of the spare bandwidth, and how such prices differ worldwide. In this work, we shed some light on this topic using dVPNs as a use-case. We start by formalizing the problem of bandwidth monetization as an optimization between a buyer's cost and seller's income. Next, we explore three popular dVPNs (Mysterium, Sentinel, and Tachyon) using both active and passive measurements. We find that dVPNs have a large and growing footprint, and offer comparable performance to their centralized counterpart. We identify Mysterium (in the US) as the most concrete realization of a bandwidth marketplace, for which we derive a value of spare Internet bandwidth ranging between 11 and 14 cents per GB. We also show that both buyers and sellers utilize ad-hoc "rules-of-thumb" when choosing their prices, which results in a sub-optimal marketplace. By applying our optimization, a seller's income can be tripled by setting a price lower than the default one which allows to attract more buyers. These observations motivate us to create RING, a first and concrete system which helps sellers to automatically adjust their prices and traffic volumes across multiple marketplaces.
{"title":"Monetizing Spare Bandwidth","authors":"Yunming Xiao, Matteo Varvello, A. Kuzmanovic","doi":"10.1145/3530899","DOIUrl":"https://doi.org/10.1145/3530899","url":null,"abstract":"Residential Internet speeds have been rapidly increasing, reaching averages of ~100 Mbps in most developed countries. Several studies have shown that users have way more bandwidth than they need, only using about 20-30% on a regular day. Several systems exploit this trend by enabling users to monetize their spare bandwidth, e.g., by sharing their WiFi connection or by participating in distributed proxy or VPN (dVPN) services. Despite the proliferation of such systems, little is known on how such marketplaces operate, what are the key factors that determine the price of the spare bandwidth, and how such prices differ worldwide. In this work, we shed some light on this topic using dVPNs as a use-case. We start by formalizing the problem of bandwidth monetization as an optimization between a buyer's cost and seller's income. Next, we explore three popular dVPNs (Mysterium, Sentinel, and Tachyon) using both active and passive measurements. We find that dVPNs have a large and growing footprint, and offer comparable performance to their centralized counterpart. We identify Mysterium (in the US) as the most concrete realization of a bandwidth marketplace, for which we derive a value of spare Internet bandwidth ranging between 11 and 14 cents per GB. We also show that both buyers and sellers utilize ad-hoc \"rules-of-thumb\" when choosing their prices, which results in a sub-optimal marketplace. By applying our optimization, a seller's income can be tripled by setting a price lower than the default one which allows to attract more buyers. These observations motivate us to create RING, a first and concrete system which helps sellers to automatically adjust their prices and traffic volumes across multiple marketplaces.","PeriodicalId":426760,"journal":{"name":"Proceedings of the ACM on Measurement and Analysis of Computing Systems","volume":"19 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2022-06-06","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"127636135","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Sheng-Chun Kao, Hyoukjun Kwon, Michael Pellauer, A. Parashar, T. Krishna
The high efficiency of domain-specific hardware accelerators for machine learning (ML) has come fromspecialization, with the trade-off of less configurability/ flexibility. There is growing interest in developingflexible ML accelerators to make them future-proof to the rapid evolution of Deep Neural Networks (DNNs). However, the notion of accelerator flexibility has always been used in an informal manner, restricting computer architects from conducting systematic apples-to-apples design-space exploration (DSE) across trillions of choices. In this work, we formally define accelerator flexibility and show how it can be integrated for DSE. % flows. Specifically, we capture DNN accelerator flexibility across four axes: %the map-space of DNN accelerator along four flexibility axes: tiling, ordering, parallelization, and array shape. We categorize existing accelerators into 16 classes based on their axes of flexibility support, and define a precise quantification of the degree of flexibility of an accelerator across each axis. We leverage these to develop a novel flexibility-aware DSE framework. %It respects the difference of accelerator flexibility classes and degree of flexibility support in different accelerators, creating unique map-spaces. %and forms a unique map space for exploration. % We demonstrate how this can be used to perform first-of-their-kind evaluations, including an isolation study to identify the individual impact of the flexibility axes. We demonstrate that adding flexibility features to a hypothetical DNN accelerator designed in 2014 improves runtime on future (i.e., present-day) DNNs by 11.8x geomean.
{"title":"A Formalism of DNN Accelerator Flexibility","authors":"Sheng-Chun Kao, Hyoukjun Kwon, Michael Pellauer, A. Parashar, T. Krishna","doi":"10.1145/3530907","DOIUrl":"https://doi.org/10.1145/3530907","url":null,"abstract":"The high efficiency of domain-specific hardware accelerators for machine learning (ML) has come fromspecialization, with the trade-off of less configurability/ flexibility. There is growing interest in developingflexible ML accelerators to make them future-proof to the rapid evolution of Deep Neural Networks (DNNs). However, the notion of accelerator flexibility has always been used in an informal manner, restricting computer architects from conducting systematic apples-to-apples design-space exploration (DSE) across trillions of choices. In this work, we formally define accelerator flexibility and show how it can be integrated for DSE. % flows. Specifically, we capture DNN accelerator flexibility across four axes: %the map-space of DNN accelerator along four flexibility axes: tiling, ordering, parallelization, and array shape. We categorize existing accelerators into 16 classes based on their axes of flexibility support, and define a precise quantification of the degree of flexibility of an accelerator across each axis. We leverage these to develop a novel flexibility-aware DSE framework. %It respects the difference of accelerator flexibility classes and degree of flexibility support in different accelerators, creating unique map-spaces. %and forms a unique map space for exploration. % We demonstrate how this can be used to perform first-of-their-kind evaluations, including an isolation study to identify the individual impact of the flexibility axes. We demonstrate that adding flexibility features to a hypothetical DNN accelerator designed in 2014 improves runtime on future (i.e., present-day) DNNs by 11.8x geomean.","PeriodicalId":426760,"journal":{"name":"Proceedings of the ACM on Measurement and Analysis of Computing Systems","volume":"53 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2022-06-06","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"130733071","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Chenxingyu Zhao, Tapan Chugh, Jaehong Min, Ming Liu, A. Krishnamurthy
LSM-tree-based key-value stores like RocksDB are widely used to support many applications. However, configuring a RocksDB instance is challenging for the following reasons: 1) RocksDB has a massive parameter space to configure; 2) there are inherent trade-offs and dependencies between parameters; 3) right configurations are dependent on workload and hardware; and 4) evaluating configurations is time-consuming. Prior works struggle with handling the curse of dimensionality, capturing relationships between parameters, adapting configurations to workload and hardware, and evaluating quickly. In this work, we present a system, Dremel, to adaptively and quickly configure RocksDB with strategies based on the Multi-Armed Bandit model. To handle the massive parameter space, we propose using fused features, which encode domain-specific knowledge, to work as a compact and powerful representation for configurations. To adapt to the workload and hardware, we build an online bandit model to identify the best configuration. To evaluate quickly, we enable multi-fidelity evaluation and upper-confidence-bound sampling to speed up identifying the best configuration. Dremel not only achieves up to ×2.61 higher IOPS and 57% less latency than default configurations but also achieves up to 63% improvements over prior works on 18 different settings with the same or less time budget.
{"title":"Dremel","authors":"Chenxingyu Zhao, Tapan Chugh, Jaehong Min, Ming Liu, A. Krishnamurthy","doi":"10.1145/3530903","DOIUrl":"https://doi.org/10.1145/3530903","url":null,"abstract":"LSM-tree-based key-value stores like RocksDB are widely used to support many applications. However, configuring a RocksDB instance is challenging for the following reasons: 1) RocksDB has a massive parameter space to configure; 2) there are inherent trade-offs and dependencies between parameters; 3) right configurations are dependent on workload and hardware; and 4) evaluating configurations is time-consuming. Prior works struggle with handling the curse of dimensionality, capturing relationships between parameters, adapting configurations to workload and hardware, and evaluating quickly. In this work, we present a system, Dremel, to adaptively and quickly configure RocksDB with strategies based on the Multi-Armed Bandit model. To handle the massive parameter space, we propose using fused features, which encode domain-specific knowledge, to work as a compact and powerful representation for configurations. To adapt to the workload and hardware, we build an online bandit model to identify the best configuration. To evaluate quickly, we enable multi-fidelity evaluation and upper-confidence-bound sampling to speed up identifying the best configuration. Dremel not only achieves up to ×2.61 higher IOPS and 57% less latency than default configurations but also achieves up to 63% improvements over prior works on 18 different settings with the same or less time budget.","PeriodicalId":426760,"journal":{"name":"Proceedings of the ACM on Measurement and Analysis of Computing Systems","volume":"23 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2022-06-06","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"133028589","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Gyeongsik Yang, C. Shin, J. Lee, Yeonho Yoo, C. Yoo
The prediction of the resource consumption for the distributed training of deep learning models is of paramount importance, as it can inform a priori users how long their training would take and also enable users to manage the cost of training. Yet, no such prediction is available for users because the resource consumption itself varies significantly according to "settings" such as GPU types and also by "workloads" like deep learning models. Previous studies have aimed to derive or model such a prediction, but they fall short of accommodating the various combinations of settings and workloads together. This study presents Driple that designs graph neural networks to predict the resource consumption of diverse workloads. Driple also designs transfer learning to extend the graph neural networks to adapt to differences in settings. The evaluation results show that Driple can effectively predict a wide range of workloads and settings. At the same time, Driple can efficiently reduce the time required to tailor the prediction for different settings by up to 7.3×.
{"title":"Prediction of the Resource Consumption of Distributed Deep Learning Systems","authors":"Gyeongsik Yang, C. Shin, J. Lee, Yeonho Yoo, C. Yoo","doi":"10.1145/3530895","DOIUrl":"https://doi.org/10.1145/3530895","url":null,"abstract":"The prediction of the resource consumption for the distributed training of deep learning models is of paramount importance, as it can inform a priori users how long their training would take and also enable users to manage the cost of training. Yet, no such prediction is available for users because the resource consumption itself varies significantly according to \"settings\" such as GPU types and also by \"workloads\" like deep learning models. Previous studies have aimed to derive or model such a prediction, but they fall short of accommodating the various combinations of settings and workloads together. This study presents Driple that designs graph neural networks to predict the resource consumption of diverse workloads. Driple also designs transfer learning to extend the graph neural networks to adapt to differences in settings. The evaluation results show that Driple can effectively predict a wide range of workloads and settings. At the same time, Driple can efficiently reduce the time required to tailor the prediction for different settings by up to 7.3×.","PeriodicalId":426760,"journal":{"name":"Proceedings of the ACM on Measurement and Analysis of Computing Systems","volume":"94 3 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2022-06-06","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"130357252","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
The ACM Proceedings of the ACM on Measurement and Analysis of Computing Systems (POMACS) focuses on the measurement and performance evaluation of computer systems and operates in close collaboration with the ACM Special Interest Group SIGMETRICS. All papers in this issue of POMACS will be presented during the ACM SIGMETRICS/Performance 2022 conference. The issue contains papers selected by the editorial board via a rigorous review process that follows a hybrid conference and journal model, with reviews conducted by the 101 members of our POMACS editorial board. Each paper was either conditionally accepted (and shepherded), allowed a "one-shot" revision (to be resubmitted to one of the subsequent two deadlines), or rejected (with resubmission allowed after a year). For this issue, which represents the winter deadline, we accepted 17 papers out of 126 submissions (including 4 papers that had been given a "one-shot" revision opportunity). All submitted papers received at least 3 reviews and we held an online TPC meeting. Based on the indicated primary track, roughly 31% of the submissions were in the Measurement & Applied Modeling track, 25% were in the Systems track, 23% were in the Theory track, and 21% were in the Learning track. Many people contributed to the success of this issue of POMACS. First, we would like to thank the authors, who submitted their work to SIGMETRICS/POMACS. Second, we would like to thank the TPC members for their work: constructive feedback in their reviews to authors, participation to online discussions and also to TPC meetings. We also thank several external reviewers who provided their expert opinion on specific submissions that required additional input. We are also grateful to the SIGMETRICS Board Chair, Giuliano Casale, and to past TPC Chairs. Finally, we are grateful to the Organization Committee and to the SIGMETRICS Board for their ongoing efforts and initiatives for creating an exciting program for ACM SIGMETRICS/Performance 2022.
{"title":"POMACS V6, N2, June 2022 Editorial","authors":"Niklas Carlsson, Edith Cohen, Philippe Robert","doi":"10.1145/3530890","DOIUrl":"https://doi.org/10.1145/3530890","url":null,"abstract":"The ACM Proceedings of the ACM on Measurement and Analysis of Computing Systems (POMACS) focuses on the measurement and performance evaluation of computer systems and operates in close collaboration with the ACM Special Interest Group SIGMETRICS. All papers in this issue of POMACS will be presented during the ACM SIGMETRICS/Performance 2022 conference. The issue contains papers selected by the editorial board via a rigorous review process that follows a hybrid conference and journal model, with reviews conducted by the 101 members of our POMACS editorial board. Each paper was either conditionally accepted (and shepherded), allowed a \"one-shot\" revision (to be resubmitted to one of the subsequent two deadlines), or rejected (with resubmission allowed after a year). For this issue, which represents the winter deadline, we accepted 17 papers out of 126 submissions (including 4 papers that had been given a \"one-shot\" revision opportunity). All submitted papers received at least 3 reviews and we held an online TPC meeting. Based on the indicated primary track, roughly 31% of the submissions were in the Measurement & Applied Modeling track, 25% were in the Systems track, 23% were in the Theory track, and 21% were in the Learning track. Many people contributed to the success of this issue of POMACS. First, we would like to thank the authors, who submitted their work to SIGMETRICS/POMACS. Second, we would like to thank the TPC members for their work: constructive feedback in their reviews to authors, participation to online discussions and also to TPC meetings. We also thank several external reviewers who provided their expert opinion on specific submissions that required additional input. We are also grateful to the SIGMETRICS Board Chair, Giuliano Casale, and to past TPC Chairs. Finally, we are grateful to the Organization Committee and to the SIGMETRICS Board for their ongoing efforts and initiatives for creating an exciting program for ACM SIGMETRICS/Performance 2022.","PeriodicalId":426760,"journal":{"name":"Proceedings of the ACM on Measurement and Analysis of Computing Systems","volume":"6 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2022-06-06","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"128853649","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Haris Bin Zia, Aravindh Raman, Ignacio Castro, Ishaku Hassan Anaobi, Emiliano De Cristofaro, Nishanth R. Sastry, Gareth Tyson
The "Decentralised Web" (DW) is an evolving concept, which encompasses technologies aimed at providing greater transparency and openness on the web. The DW relies on independent servers (aka instances) that mesh together in a peer-to-peer fashion to deliver a range of services (e.g. micro-blogs, image sharing, video streaming). However, toxic content moderation in this decentralised context is challenging. This is because there is no central entity that can define toxicity, nor a large central pool of data that can be used to build universal classifiers. It is therefore unsurprising that there have been several high-profile cases of the DW being misused to coordinate and disseminate harmful material. Using a dataset of 9.9M posts from 117K users on Pleroma (a popular DW microblogging service), we quantify the presence of toxic content. We find that toxic content is prevalent and spreads rapidly between instances. We show that automating per-instance content moderation is challenging due to the lack of sufficient training data available and the effort required in labelling. We therefore propose and evaluate ModPair, a model sharing system that effectively detects toxic content, gaining an average per-instance macro-F1 score 0.89.
{"title":"Toxicity in the Decentralized Web and the Potential for Model Sharing","authors":"Haris Bin Zia, Aravindh Raman, Ignacio Castro, Ishaku Hassan Anaobi, Emiliano De Cristofaro, Nishanth R. Sastry, Gareth Tyson","doi":"10.1145/3530901","DOIUrl":"https://doi.org/10.1145/3530901","url":null,"abstract":"The \"Decentralised Web\" (DW) is an evolving concept, which encompasses technologies aimed at providing greater transparency and openness on the web. The DW relies on independent servers (aka instances) that mesh together in a peer-to-peer fashion to deliver a range of services (e.g. micro-blogs, image sharing, video streaming). However, toxic content moderation in this decentralised context is challenging. This is because there is no central entity that can define toxicity, nor a large central pool of data that can be used to build universal classifiers. It is therefore unsurprising that there have been several high-profile cases of the DW being misused to coordinate and disseminate harmful material. Using a dataset of 9.9M posts from 117K users on Pleroma (a popular DW microblogging service), we quantify the presence of toxic content. We find that toxic content is prevalent and spreads rapidly between instances. We show that automating per-instance content moderation is challenging due to the lack of sufficient training data available and the effort required in labelling. We therefore propose and evaluate ModPair, a model sharing system that effectively detects toxic content, gaining an average per-instance macro-F1 score 0.89.","PeriodicalId":426760,"journal":{"name":"Proceedings of the ACM on Measurement and Analysis of Computing Systems","volume":"20 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"2022-06-06","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"121294744","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","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：