The aggressive scaling of technology may have helped to meet the growing demand for higher memory capacity and density, but has also made DRAM cells more prone to errors. Such a reality triggered a lot of interest in modeling DRAM behavior for either predicting the errors in advance or for adjusting DRAM circuit parameters to achieve a better trade-off between energy efficiency and reliability. Existing modeling efforts may have studied the impact of few operating parameters and temperature on DRAM reliability using custom FPGAs setups, however they neglected the combined effect of workload-specific features that can be systematically investigated only on a real system. In this paper, we present the results of our study on workload-dependent DRAM error behavior within a real server considering various operating parameters, such as the refresh rate, voltage and temperature. We show that the rate of single- and multi-bit errors may vary across workloads by 8x, indicating that program inherent features can affect DRAM reliability significantly. Based on this observation, we extract 249 features, such as the memory access rate, the rate of cache misses, the memory reuse time and data entropy, from various compute-intensive, caching and analytics benchmarks. We apply several supervised learning methods to construct the DRAM error behavior model for 72 server-grade DRAM chips using the memory operating parameters and extracted program inherent features. Our results show that, with an appropriate choice of program features and supervised learning method, the rate of single- and multi-bit errors can be predicted for a specific DRAM module with an average error of less than 10.5 %, as opposed to the 2.9x estimation error obtained for a conventional workload-unaware error model.
As the cost-per-byte of storage systems dramatically decreases, SSDs are finding their ways in emerging cloud infrastructure. Similar trend is happening for main memory subsystem, as advanced DRAM technologies with higher capacity, frequency and number of channels are deploying for cloud-scale solutions specially for non-virtualized environment where cloud subscribers can exactly specify the configuration of underling hardware. Given the performance sensitivity of standard workloads to the memory hierarchy parameters, it is important to understand the role of memory and storage for data intens…
Major chip manufacturers have all introduced multicore microprocessors. Multi-socket systems built from these processors are used for running various server applications. However to the best of our knowledge current commercial operating systems are not optimized for multi-threaded workloads running on such servers. Cache-to-cache transfers and remote memory accesses impact the performance of such workloads. This paper presents a unified approach to optimizing OS scheduling algorithms for both cache-to-cache transfers and remote DRAM accesses that also takes cache affinity into account. By obse…
Multi-socket multi-core servers are used for solving some of the important problems in computing. Remote DRAM accesses can impact performance of certain applications running on such servers. This paper presents a new near linear operating system (OS) scheduling algorithm to reduce the impact of such remote DRAM accesses. By keeping track of the number of local and remote DRAM accesses, using performance counters, for every thread and applying this algorithm, I come up with a new schedule of threads for the next quantum. This new schedule reduces remote DRAM accesses and improves overall perfor…
Major chip manufacturers have all introduced multicore microprocessors. Multi-socket systems built from these processors are used for running various server applications. Depending on the application that is run on the system, remote memory accesses can impact overall performance. This paper presents a cache optimization that can cut down remote DRAM accesses. By keeping track of remote cache lines loaded from remote DRAM and by biasing the cache replacement policy towards such remote DRAM cache lines the number of cache misses are reduced. This in turn results in improvement of overall perfor…
Major chip manufacturers have all introduced multicore microprocessors. Multi-socket systems built from these processors are routinely used for running various server applications. Depending on the application that is run on the system, remote memory accesses can impact overall performance. This paper presents a new operating system (OS) scheduling optimization to reduce the impact of such remote memory accesses. By observing the pattern of local and remote DRAM accesses for every thread in each scheduling quantum and applying different algorithms, we come up with a new schedule of threads for…
SWARM+: Scalable and Resilient Multi-Agent Consensus for Decentralized Data-Aware Workload Management
Renaissance Computing Institute, University of North Carolina at Chapel Hill（北卡罗来纳大学教堂山分校文艺复兴计算研究所）；Argonne National Laboratory（阿贡国家实验室）；Information Sciences Institute, University of Southern California（南加州大学信息科学研究所）
Deep neural networks (DNNs) have emerged as successful solutions for variety of artificial intelligence applications, but their very large and deep models impose high computational requirements during training. Multi-GPU parallelization is a popular option to accelerate demanding computations in DNN training, but most state-of-the-art multi-GPU deep learning frameworks not only require users to have an in-depth understanding of the implementation of the frameworks themselves, but also apply parallelization in a straight-forward way without optimizing GPU utilization. In this work, we propose a…
Multi-server jobs are imperative in modern computing clusters. A multi-server job has multiple task components and each of the task components is responsible for processing a specific size of workloads. Efficient online workload dispatching is crucial but challenging to co-located heterogeneous multi-server jobs. The dispatching policy should decide $(i)$ where to launch each task component instance of the arrived jobs and $(ii)$ the size of workloads that each task component processes. Existing policies are explicit and effective when facing service locality and resource contention in both of…
Database platform-as-a-service (dbPaaS) is developing rapidly and a large number of databases have been migrated to run on the Clouds for the low cost and flexibility. Emerging Clouds rely on the tenants to provide the resource specification for their database workloads. However, they tend to over-estimate the resource requirement of their databases, resulting in the unnecessarily high cost and low Cloud utilization. A methodology that automatically suggests the "just-enough" resource specification that fulfills the performance requirement of every database workload is profitable. To this end,…
Data center providers seek to minimize their total cost of ownership (TCO), while power consumption has become a social concern. We present formulations to minimize server energy consumption and server cost under three different data center server setups (homogeneous, heterogeneous, and hybrid hetero-homogeneous clusters) with dynamic temporal workload. Our studies show that the homogeneous model significantly differs from the heterogeneous model in computational time (by an order of magnitude). To be able to compute optimal configurations in near real-time for large scale data centers, we pro…
Cloud service providers are distributing data centers geographically to minimize energy costs through intelligent workload distribution. With increasing data volumes in emerging cloud workloads, it is critical to factor in the network costs for transferring workloads across data centers. For geo-distributed data centers, many researchers have been exploring strategies for energy cost minimization and intelligent inter-data-center workload distribution separately. However, prior work does not comprehensively and simultaneously consider data center energy costs, data transfer costs, and data cen…
Autoscaling is a critical component for efficient resource utilization with satisfactory quality of service (QoS) in cloud computing. This paper investigates proactive autoscaling for widely-used scaling-per-query applications where scaling is required for each query, such as container registry and function-as-a-service (FaaS). In these scenarios, the workload often exhibits high uncertainty with complex temporal patterns like periodicity, noises and outliers. Conservative strategies that scale out unnecessarily many instances lead to high resource costs whereas aggressive strategies may resul…
Specialized accelerators such as GPUs, TPUs, FPGAs, and custom ASICs have been increasingly deployed to train deep learning models. These accelerators exhibit heterogeneous performance behavior across model architectures. Existing schedulers for clusters of accelerators, which are used to arbitrate these expensive training resources across many users, have shown how to optimize for various multi-job, multi-user objectives, like fairness and makespan. Unfortunately, existing schedulers largely do not consider performance heterogeneity. In this paper, we propose Gavel, a heterogeneity-aware sche…
Distributed Stream Processing systems have become an essential part of big data processing platforms. They are characterized by the high-throughput processing of near to real-time event streams with the goal of delivering low-latency results and thus enabling time-sensitive decision making. At the same time, results are expected to be consistent even in the presence of partial failures where exactly-once processing guarantees are required for correctness. Stream processing workloads are oftentimes dynamic in nature which makes static configurations highly inefficient as time goes by. Static re…
Byte-addressable non-volatile memory (NVM) features high density, DRAM comparable performance, and persistence. These characteristics position NVM as a promising new tier in the memory hierarchy. Nevertheless, NVM has asymmetric read and write performance, and considerably higher write energy than DRAM. Our work provides an in-depth evaluation of the first commercially available byte-addressable NVM -- the Intel Optane DC persistent memory. The first part of our study quantifies the latency, bandwidth, power efficiency, and energy consumption under eight memory configurations. We also evaluate…
New HPC machines are getting close to the exascale. Power consumption for those machines has been increasing, and researchers are studying ways to reduce it. A second trend is HPC machines' growing complexity, with increasing heterogeneous hardware components and different clusters architectures cooperating in the same machine. We refer to these environments with the term heterogeneous multi-cluster environments. With the aim of optimizing performance and energy consumption in these environments, this paper proposes an Energy-Aware-Multi-Cluster (EAMC) job scheduling policy. EAMC-policy is abl…
Large-scale datasets in the form of knowledge graphs are often used in numerous domains, today. A knowledge graphs size often exceeds the capacity of a single computer system, especially if the graph must be stored in main memory. To overcome this, knowledge graphs can be partitioned into multiple sub-graphs and distributed as shards among many computing nodes. However, performance of many common tasks performed on graphs, such as querying, suffers, as a result. This is due to distributed joins mandated by graph edges crossing (cutting) the partitions. In this paper, we propose a method of kno…
EchoFlow: A Workload-Aware Parameter Tuning Method for Blockchain Systems
Institute of Computing Technology, Chinese Academy of Sciences（中国科学院计算技术研究所）；School of Computer Science and Technology, UCAS（中国科学院大学计算机科学与技术学院）；Hangzhou Qulian Technology Co., Ltd.（杭州趣链科技有限公司）；State Key Laboratory of Blockchain and Data Security, ZJU（浙江大学区块链与数据安全全国重点实验室）；Beijing Advanced Innovation Center for Future Blockchain and Privacy Computing, Beihang University（北京航空航天大学未来区块链与隐私计算先进创新中心）；Shandong Key Laboratory Blockchain Finance（山东省区块链金融重点实验室）
When processing data streams with highly skewed and nonstationary key distributions, we often observe overloaded partitions when the hash partitioning fails to balance data correctly. To avoid slow tasks that delay the completion of the whole stage of computation, it is necessary to apply adaptive, on-the-fly partitioning that continuously recomputes an optimal partitioner, given the observed key distribution. While such solutions exist for batch processing of static data sets and stateless stream processing, the task is difficult for long-running stateful streaming jobs where key distribution…
The growing electricity demand of cloud and edge computing increases operational costs and will soon have a considerable impact on the environment. A possible countermeasure is equipping IT infrastructure directly with on-site renewable energy sources. Yet, particularly smaller data centers may not be able to use all generated power directly at all times, while feeding it into the public grid or energy storage is often not an option. To maximize the usage of renewable excess energy, we propose Cucumber, an admission control policy that accepts delay-tolerant workloads only if they can be compu…
Large-scale knowledge graphs are increasingly common in many domains. Their large sizes often exceed the limits of systems storing the graphs in a centralized data store, especially if placed in main memory. To overcome this, large knowledge graphs need to be partitioned into multiple sub-graphs and placed in nodes in a distributed system. But querying these fragmented sub-graphs poses new challenges, such as increased communication costs, due to distributed joins involving cut edges. To combat these problems, a good partitioning should reduce the edge cuts while considering a given query work…
In modern networking research, infrastructure-assisted unmanned autonomous vehicles (UAVs) are actively considered for real-time learning-based surveillance and aerial data-delivery under unexpected 3D free mobility and coordination. In this system model, it is essential to consider the power limitation in UAVs and autonomous object recognition (for abnormal behavior detection) deep learning performance in infrastructure/towers. To overcome the power limitation of UAVs, this paper proposes a novel aerial scheduling algorithm between multi-UAVs and multi-towers where the towers conduct wireless…
We present Recipe, a principled approach for converting concurrent DRAM indexes into crash-consistent indexes for persistent memory (PM). The main insight behind Recipe is that isolation provided by a certain class of concurrent in-memory indexes can be translated with small changes to crash-consistency when the same index is used in PM. We present a set of conditions that enable the identification of this class of DRAM indexes, and the actions to be taken to convert each index to be persistent. Based on these conditions and conversion actions, we modify five different DRAM indexes based on B+…
University of Washington（华盛顿大学）；Microsoft Research（微软研究院）；Facebook AI（脸书人工智能研究院）
ML workloads are becoming increasingly popular in the cloud. Good cloud training performance is contingent on efficient parameter exchange among VMs. We find that Collectives, the widely used distributed communication algorithms, cannot perform optimally out of the box due to the hierarchical topology of datacenter networks and multi-tenancy nature of the cloudenvironment.In this paper, we present Cloud Collectives , a prototype that accelerates collectives by reordering theranks of participating VMs such that the communication pattern dictated by the selected collectives operation best exploi…
DALI: A Workload-Aware Offloading Framework for Efficient MoE Inference on Local PCs
Institute of Automation, CAS（中国科学院自动化研究所）；School of Future Technology, University of Chinese Academy of Sciences（中国科学院大学未来技术学院）；Shanghai Jiao Tong University（上海交通大学）；AiRiA
GPUs are essential to accelerating the latency-sensitive deep neural network (DNN) inference workloads in cloud datacenters. To fully utilize GPU resources, spatial sharing of GPUs among co-located DNN inference workloads becomes increasingly compelling. However, GPU sharing inevitably brings severe performance interference among co-located inference workloads, as motivated by an empirical measurement study of DNN inference on EC2 GPU instances. While existing works on guaranteeing inference performance service level objectives (SLOs) focus on either temporal sharing of GPUs or reactive GPU re…
Datacenters suffer from resource utilization inefficiencies due to the conflicting goals of service owners and platform providers. Service owners intending to maintain Service Level Objectives (SLO) for themselves typically request a conservative amount of resources. Platform providers want to increase operational efficiency to reduce capital and operating costs. Achieving both operational efficiency and SLO for individual services at the same time is challenging due to the diversity in service workload characteristics, resource usage patterns that are dependent on input load, heterogeneity in…
Data intensive applications often involve the analysis of large datasets that require large amounts of compute and storage resources. While dedicated compute and/or storage farms offer good task/data throughput, they suffer low resource utilization problem under varying workloads conditions. If we instead move such data to distributed computing resources, then we incur expensive data transfer cost. In this paper, we propose a data diffusion approach that combines dynamic resource provisioning, on-demand data replication and caching, and data locality-aware scheduling to achieve improved resour…
We present the Cuckoo Trie, a fast, memory-efficient ordered index structure. The Cuckoo Trie is designed to have memory-level parallelism -- which a modern out-of-order processor can exploit to execute DRAM accesses in parallel -- without sacrificing memory efficiency. The Cuckoo Trie thus breaks a fundamental performance barrier faced by current indexes, whose bottleneck is a series of dependent pointer-chasing DRAM accesses -- e.g., traversing a search tree path -- which the processor cannot parallelize. Our evaluation shows that the Cuckoo Trie outperforms state-of-the-art-indexes by up to…
GreenDirector: carbon- and water-aware workload placement for sustainable computing
Instituto de Física de Cantabria (IFCA), CSIC-UC（坎塔布里亚物理研究所 (IFCA)，西班牙国家研究委员会-坎塔布里亚大学）；Aix Marseille Univ, CNRS/IN2P3, CPPM（艾克斯马赛大学，法国国家科学研究中心/法国国家粒子物理研究所，普罗旺斯粒子与核物理中心）
Massive exploitation of next-generation sequencing technologies requires dealing with both: huge amounts of data and complex bioinformatics pipelines. Computing architectures have evolved to deal with these problems, enabling approaches that were unfeasible years ago: accelerators and Non-Volatile Memories (NVM) are becoming widely used to enhance the most demanding workloads. However, bioinformatics workloads are usually part of bigger pipelines with different and dynamic needs in terms of resources. The introduction of Software Defined Infrastructures (SDI) for data centers provides roots to…
Today's clusters often have to divide resources among a diverse set of jobs. These jobs are heterogeneous both in execution time and in their rate of arrival. Execution time heterogeneity has lead to the development of hybrid schedulers that can schedule both short and long jobs to ensure good task placement. However, arrival rate heterogeneity, or burstiness, remains a problem in existing schedulers. These hybrid schedulers manage resources on statically provisioned cluster, which can quickly be overwhelmed by bursts in the number of arriving jobs. In this paper we propose CloudCoaster, a hyb…
