Academic literature on the topic 'Memory disaggregation'

Hardware disaggregation has emerged as one of the most fundamental shifts in how we build computer systems over the past decades. While disaggregation has been successful for several types of resources (storage, power, and others), memory disaggregation has yet to happen. We make the case that the time for memory disaggregation has arrived. We look at past successful disaggregation stories and learn that their success depended on two requirements: addressing a burning issue and being technically feasible. We examine memory disaggregation through this lens and find that both requirements are finally met. Once available, memory disaggregation will require software support to be used effectively. We discuss some of the challenges of designing an operating system that can utilize disaggregated memory for itself and its applications.

While permissioned blockchains enable a family of data center applications, existing systems suffer from imbalanced loads across compute and memory, exacerbating the underutilization of cloud resources. This paper presents FlexChain , a novel permissioned blockchain system that addresses this challenge by physically disaggregating CPUs, DRAM, and storage devices to process different blockchain workloads efficiently. Disaggregation allows blockchain service providers to upgrade and expand hardware resources independently to support a wide range of smart contracts with diverse CPU and memory demands. Moreover, it ensures efficient resource utilization and hence prevents resource fragmentation in a data center. We have explored the design of XOV blockchain systems in a disaggregated fashion and developed a tiered key-value store that can elastically scale its memory and storage. Our design significantly speeds up the execution stage. We have also leveraged several techniques to parallelize the validation stage in FlexChain to further improve the overall blockchain performance. Our evaluation results show that FlexChain can provide independent compute and memory scalability, while incurring at most 12.8% disaggregation overhead. FlexChain achieves almost identical throughput as the state-of-the-art distributed approaches with significantly lower memory and CPU consumption for compute-intensive and memory-intensive workloads respectively.

Compute and memory are tightly coupled within each server in traditional datacenters. Large-scale datacenter operators have identified this coupling as a root cause behind fleetwide resource underutilization and increasing Total Cost of Ownership (TCO). With the advent of ultra-fast networks and cache-coherent interfaces, memory disaggregation has emerged as a potential solution, whereby applications can leverage available memory even outside server boundaries. This paper summarizes the growing research landscape of memory disaggregation from a software perspective and introduces the challenges toward making it practical under current and future hardware trends. We also reflect on our seven-year journey in the SymbioticLab to build a comprehensive disaggregated memory system over ultra-fast networks. We conclude with some open challenges toward building next-generation memory disaggregation systems leveraging emerging cache-coherent interconnects.

Large-scale aggregation and its inverse, disaggregation, problems are important in many fields of studies like macroeconomics, astronomy, hydrology and sociology. It was shown in Granger (1980) that a certain aggregation of random coefficient AR(1) models can lead to long memory output. Dacunha-Castelle and Oppenheim (2001) explored the topic further, answering when and if a predefined long memory process could be obtained as the result of aggregation of a specific class of individual processes. In this paper, the disaggregation scheme of Leipus et al. (2006) is briefly discussed. Then disaggregation into AR(1) is analyzed further, resulting in a theorem that helps, under corresponding assumptions, to construct a mixture density for a given aggregated by AR(1) scheme process. Finally the theorem is illustrated by FARUMA mixture densityÆs example.

Memory disaggregation is a promising architecture for modern datacenters that separates compute and memory resources into independent pools connected by ultra-fast networks, which can improve memory utilization, reduce cost, and enable elastic scaling of compute and memory resources. However, existing memory disaggregation solutions based on remote direct memory access (RDMA) suffer from high latency and additional overheads including page faults and code refactoring. Emerging cache-coherent interconnects such as CXL offer opportunities to reconstruct high-performance memory disaggregation. However, existing CXL-based approaches have physical distance limitation and cannot be deployed across racks. In this article, we propose Rcmp, a novel low-latency and highly scalable memory disaggregation system based on RDMA and CXL. The significant feature is that Rcmp improves the performance of RDMA-based systems via CXL, and leverages RDMA to overcome CXL’s distance limitation. To address the challenges of the mismatch between RDMA and CXL in terms of granularity, communication, and performance, Rcmp (1) provides a global page-based memory space management and enables fine-grained data access, (2) designs an efficient communication mechanism to avoid communication blocking issues, (3) proposes a hot-page identification and swapping strategy to reduce RDMA communications, and (4) designs an RDMA-optimized RPC framework to accelerate RDMA transfers. We implement a prototype of Rcmp and evaluate its performance by using micro-benchmarks and running a key-value store with YCSB benchmarks. The results show that Rcmp can achieve 5.2× lower latency and 3.8× higher throughput than RDMA-based systems. We also demonstrate that Rcmp can scale well with the increasing number of nodes without compromising performance.

It is challenging for cloud-native relational databases to meet the ever-increasing needs of scaling compute and memory resources independently and elastically. The recent emergence of memory disaggregation architecture, relying on high-speed RDMA network, offers opportunities to build cost-effective and elastic cloud-native databases. There exist proposals to let unmodified applications run transparently on disaggregated systems. However, running relational database kernel atop such proposals experiences notable performance degradation and time-consuming failure recovery, offsetting the benefits of disaggregation. To address these challenges, in this paper, we propose a novel database architecture called LegoBase, which explores the co-design of database kernel and memory disaggregation. It pushes the memory management back to the database layer for bypassing the Linux I/O stack and re-using or designing (remote) memory access optimizations with an understanding of data access patterns. LegoBase further splits the conventional ARIES fault tolerance protocol to independently handle the local and remote memory failures for fast recovery of compute instances. We implemented LegoBase atop MySQL. We compare LegoBase against MySQL running on a standalone machine and the state-of-the-art disaggregation proposal Infiniswap. Our evaluation shows that even with a large fraction of data placed on the remote memory, LegoBase's system performance in terms of throughput (up to 9.41% drop) and P99 latency (up to 11.58% increase) is comparable to the monolithic MySQL setup, and significantly outperforms (1.99x-2.33x, respectively) the deployment of MySQL over Infiniswap. Meanwhile, LegoBase introduces an up to 3.87x and 5.48x speedup of the recovery and warm-up time, respectively, over the monolithic MySQL and MySQL over Infiniswap, when handling failures or planned re-configurations.

Memory disaggregation (MD) allows for scalable and elastic data center design by separating compute (CPU) from memory. With MD, compute and memory are no longer coupled into the same server box. Instead, they are connected to each other via ultra-fast networking such as RDMA. MD can bring many advantages, e.g., higher memory utilization, better independent scaling (of compute and memory), and lower cost of ownership. This paper makes the case that MD can fuel the next wave of innovation on database systems. We observe that MD revives the great debate of "shared what" in the database community. We envision that distributed shared-memory databases (DSM-DB, for short) - that have not received much attention before - can be promising in the future with MD. We present a list of challenges and opportunities that can inspire next steps in system design making the case for DSM-DB.

Les ordinateurs modernes sont construits autour de deux éléments : leur CPU etleur mémoire principale volatile, ou RAM. Depuis les années 1970, ce principe a étéconstamment amélioré pour offrir toujours plus de fonctionnalités et de performances.Dans cette thèse, nous étudions deux paradigmes de mémoire qui proposent denouvelles façons d'interagir avec la mémoire dans les systèmes modernes : la mémoirenon-volatile et les accès mémoire distants. Nous mettons en œuvre des outils logicielsqui exploitent ces nouvelles approches afin de les rendre compatibles et d'exploiterleurs performances avec des applications concrètes. Nous analysons égalementl'impact des technologies utilisées, et les perspectives de leur évolution dans lesannées à venir.Pour la mémoire non-volatile, comme les performances de la mémoire sont essentiellespour atteindre le potentiel d'un CPU, cette fonctionnalité a historiquement été abandonnée.Même si les premiers ordinateurs ont été conçus avec des formes de mémoire nonvolatiles, les architectes informatiques ont commencé à utiliser la RAM volatilepour ses performances inégalées, et n'ont jamais remis en question cette décisionpendant des années. Cependant, en 2019, Intel a commercialisé un nouveau composantappelé Optane DCPMM qui rend possible l'utilisation de NVMM. Ce produit proposeune nouvelle façon de penser la persistance des données. Mais il remet égalementen question l'architecture de nos machines et la manière dont nous les programmons.Avec cette nouvelle forme de mémoire, nous avons implémenté NVCACHE, un cacheen mémoire non-volatile qui permet d'accélérer les interactions avec des supportsde stockage persistants plus lents, tels que les SSD. Nous montrons que NVCACHEest particulièrement performant pour les tâches qui nécessitent une granularitéélevée des garanties de persistance, tout en étant aussi simple à utiliser que l'interfacePOSIX traditionnelle. Comparé aux systèmes de fichiers conçus pour NVMM, NVCACHEpeut atteindre un débit similaire ou supérieur lorsque la mémoire non volatile estutilisée. De plus, NVCACHE permet aux programmes d'exploiter les performancesde NVMM sans être limité par la quantité de NVMM installée sur la machine.Un autre changement majeur dans le paysage informatique a été la popularité dessystèmes distribués. Alors que les machines ont individuellement tendance à atteindredes limites de performances, l'utilisation de plusieurs machines et le partage destâches sont devenus la nouvelle façon de créer des ordinateurs puissants. Bien quece mode de calcul permette d'augmenter le nombre de CPU utilisés simultanément,il nécessite une connexion rapide entre les nœuds de calcul. Pour cette raison,plusieurs protocoles de communication ont implémententé RDMA, un moyen delire ou d'écrire directement dans la mémoire d'un serveur distant. RDMA offre defaibles latences et un débit élevé, contournant de nombreuses étapes de la pileréseau.Cependant, RDMA reste limité dans ses fonctionnalités natives. Par exemple, iln'existe pas d'équivalent de multicast pour les fonctions RDMA les plus efficaces.Grâce à un switch programmable (le switch Intel Tofino), nous avons implémentéun mode spécial pour RDMA qui permet de lire ou d'écrire sur plusieurs serveursen même temps, sans pénalité de performances. Notre système appelé Byp4ss faitparticiper le switch aux transferts, en dupliquant les paquets RDMA. Grâce à Byp4ss,nous avons implémenté un protocole de consensus nommé DISMU. De par sa conception,DISMU est optimal en termes de latence et de débit, car il peut réduire au minimumle nombre de paquets échangés sur le réseau pour parvenir à un consensus.Enfin, en utilisant ces deux technologies, nous remarquons que les futures générationsde matériel pourraient nécessiter une nouvelle interface pour les mémoires detoutes sortes, afin de faciliter l'interopérabilité dans des systèmes qui ont tendanceà devenir de plus en plus hétérogènes et complexes
Modern computers are built around two main parts: their Central Processing Unit (CPU), and their volatile main memory, or Random Access Memory (RAM). The basis of this architecture takes its roots in the 1970's first computers. Since, this principle has been constantly upgraded to provide more functionnality and performance.In this thesis, we study two memory paradigms that drastically change the way we can interact with memory in modern systems: non-volatile memory and remote memory access. We implement software tools that leverage them in order to make them compatible and exploit their performance with concrete applications. We also analyze the impact of the technologies underlying these new memory medium, and the perspectives of their evolution in the coming years.For non-volatile memory, as the main memory performance is key to unlock the full potential of a CPU, this feature has historically been abandoned on the race for performance. Even if the first computers were designed with non-volatile forms of memory, computer architects started to use volatile RAM for its incomparable performance compared to durable storage, and never questioned this decision for years. However, in 2019 Intel released a new component called Optane DC Persistent Memory (DCPMM), a device that made possible the use of Non-Volatile Main Memory (NVMM). That product, by its capabilities, provides a new way of thinking about data persistence. Yet, it also challenges the hardware architecture used in our current machines and the way we program them.With this new form of memory we implemented NVCACHE, a cache designed for non-volatile memory that helps boosting the interactions with slower persistent storage medias, such as solid state drive (SSD). We find NVCACHE to be quite performant for workloads that require a high granularity of persistence guarantees, while being as easy to use as the traditional POSIX interface. Compared to file systems designed for NVMM, NVCACHE can reach similar or higher throughput when the non-volatile memory is used. In addition, NVCACHE allows the code to exploit NVMM performance while not being limited by the amount of NVMM installed in the machine.Another major change of in the computer landscape has been the popularity of distributed systems. As individual machines tend to reach performance limitations, using several machines and sharing workloads became the new way to build powerful computers. While this mode of computation allows the software to scale up the number of CPUs used simultaneously, it requires fast interconnection between the computing nodes. For that reason, several communication protocols implemented Remote Direct Memory Access (RDMA), a way to read or write directly into a distant machine's memory. RDMA provides low latencies and high throughput, bypassing many steps of the traditional network stack.However, RDMA remains limited in its native features. For instance, there is no advanced multicast equivalent for the most efficient RDMA functions. Thanks to a programmable switch (the Intel Tofino), we implemented a special mode for RDMA that allows a client to read or write in multiple servers at the same time, with no performance penalty. Our system called Byp4ss makes the switch participate in transfers, duplicating RDMA packets. On top of Byp4ss, we implement a consensus protocol named DISMU, which shows the typical use of Byp4ss features and its impact on performance. By design, DISMU is optimal in terms of latency and throughput, as it can reduce to the minimum the number of packets exchanged through the network to reach a consensus.Finally, by using these two technologies, we notice that future generations of hardware may require a new interface for memories of all kinds, in order to ease the interoperability in systems that tend to get more and more heterogeneous and complex

In recent years there has been an extraordinary growth of the Internet of Things (IoT) and its protocols. The increasing diffusion of electronic devices with identification, computing and communication capabilities is laying ground for the emergence of a highly distributed service and networking environment. The above mentioned situation implies that there is an increasing demand for advanced IoT data management and processing platforms. Such platforms require support for multiple protocols at the edge for extended connectivity with the objects, but also need to exhibit uniform internal data organization and advanced data processing capabilities to fulfill the demands of the application and services that consume IoT data. One of the initial approaches to address this demand is the integration between IoT and the Cloud computing paradigm. There are many benefits of integrating IoT with Cloud computing. The IoT generates massive amounts of data, and Cloud computing provides a pathway for that data to travel to its destination. But today’s Cloud computing models do not quite fit for the volume, variety, and velocity of data that the IoT generates. Among the new technologies emerging around the Internet of Things to provide a new whole scenario, the Fog Computing paradigm has become the most relevant. Fog computing was introduced a few years ago in response to challenges posed by many IoT applications, including requirements such as very low latency, real-time operation, large geo-distribution, and mobility. Also this low latency, geo-distributed and mobility environments are covered by the network architecture MEC (Mobile Edge Computing) that provides an IT service environment and Cloud-computing capabilities at the edge of the mobile network, within the Radio Access Network (RAN) and in close proximity to mobile subscribers. Fog computing addresses use cases with requirements far beyond Cloud-only solution capabilities. The interplay between Cloud and Fog computing is crucial for the evolution of the so-called IoT, but the reach and specification of such interplay is an open problem. This thesis aims to find the right techniques and design decisions to build a scalable distributed system for the IoT under the Fog Computing paradigm to ingest and process data. The final goal is to explore the trade-offs and challenges in the design of a solution from Edge to Cloud to address opportunities that current and future technologies will bring in an integrated way. This thesis describes an architectural approach that addresses some of the technical challenges behind the convergence between IoT, Cloud and Fog with special focus on bridging the gap between Cloud and Fog. To that end, new models and techniques are introduced in order to explore solutions for IoT environments. This thesis contributes to the architectural proposals for IoT ingestion and data processing by 1) proposing the characterization of a platform for hosting IoT workloads in the Cloud providing multi-tenant data stream processing capabilities, the interfaces over an advanced data-centric technology, including the building of a state-of-the-art infrastructure to evaluate the performance and to validate the proposed solution. 2) studying an architectural approach following the Fog paradigm that addresses some of the technical challenges found in the first contribution. The idea is to study an extension of the model that addresses some of the central challenges behind the converge of Fog and IoT. 3) Design a distributed and scalable platform to perform IoT operations in a moving data environment. The idea after study data processing in Cloud, and after study the convenience of the Fog paradigm to solve the IoT close to the Edge challenges, is to define the protocols, the interfaces and the data management to solve the ingestion and processing of data in a distributed and orchestrated manner for the Fog Computing paradigm for IoT in a moving data environment.
En els últims anys hi ha hagut un gran creixement del Internet of Things (IoT) i els seus protocols. La creixent difusió de dispositius electrònics amb capacitats d'identificació, computació i comunicació esta establint les bases de l’aparició de serveis altament distribuïts i del seu entorn de xarxa. L’esmentada situació implica que hi ha una creixent demanda de plataformes de processament i gestió avançada de dades per IoT. Aquestes plataformes requereixen suport per a múltiples protocols al Edge per connectivitat amb el objectes, però també necessiten d’una organització de dades interna i capacitats avançades de processament de dades per satisfer les demandes de les aplicacions i els serveis que consumeixen dades IoT. Una de les aproximacions inicials per abordar aquesta demanda és la integració entre IoT i el paradigma del Cloud computing. Hi ha molts avantatges d'integrar IoT amb el Cloud. IoT genera quantitats massives de dades i el Cloud proporciona una via perquè aquestes dades viatgin a la seva destinació. Però els models actuals del Cloud no s'ajusten del tot al volum, varietat i velocitat de les dades que genera l'IoT. Entre les noves tecnologies que sorgeixen al voltant del IoT per proporcionar un escenari nou, el paradigma del Fog Computing s'ha convertit en la més rellevant. Fog Computing es va introduir fa uns anys com a resposta als desafiaments que plantegen moltes aplicacions IoT, incloent requisits com baixa latència, operacions en temps real, distribució geogràfica extensa i mobilitat. També aquest entorn està cobert per l'arquitectura de xarxa MEC (Mobile Edge Computing) que proporciona serveis de TI i capacitats Cloud al edge per la xarxa mòbil dins la Radio Access Network (RAN) i a prop dels subscriptors mòbils. El Fog aborda casos d’us amb requisits que van més enllà de les capacitats de solucions només Cloud. La interacció entre Cloud i Fog és crucial per a l'evolució de l'anomenat IoT, però l'abast i especificació d'aquesta interacció és un problema obert. Aquesta tesi té com objectiu trobar les decisions de disseny i les tècniques adequades per construir un sistema distribuït escalable per IoT sota el paradigma del Fog Computing per a ingerir i processar dades. L'objectiu final és explorar els avantatges/desavantatges i els desafiaments en el disseny d'una solució des del Edge al Cloud per abordar les oportunitats que les tecnologies actuals i futures portaran d'una manera integrada. Aquesta tesi descriu un enfocament arquitectònic que aborda alguns dels reptes tècnics que hi ha darrere de la convergència entre IoT, Cloud i Fog amb especial atenció a reduir la bretxa entre el Cloud i el Fog. Amb aquesta finalitat, s'introdueixen nous models i tècniques per explorar solucions per entorns IoT. Aquesta tesi contribueix a les propostes arquitectòniques per a la ingesta i el processament de dades IoT mitjançant 1) proposant la caracterització d'una plataforma per a l'allotjament de workloads IoT en el Cloud que proporcioni capacitats de processament de flux de dades multi-tenant, les interfícies a través d'una tecnologia centrada en dades incloent la construcció d'una infraestructura avançada per avaluar el rendiment i validar la solució proposada. 2) estudiar un enfocament arquitectònic seguint el paradigma Fog que aborda alguns dels reptes tècnics que es troben en la primera contribució. La idea és estudiar una extensió del model que abordi alguns dels reptes centrals que hi ha darrere de la convergència de Fog i IoT. 3) Dissenyar una plataforma distribuïda i escalable per a realitzar operacions IoT en un entorn de dades en moviment. La idea després d'estudiar el processament de dades en el Cloud, i després d'estudiar la conveniència del paradigma Fog per resoldre els desafiaments de IoT a prop del Edge, és definir els protocols, les interfícies i la gestió de dades per resoldre la ingestió i processament de dades d’una manera més eficient

碩士
國立臺灣科技大學
電機工程系
107
Non-intrusive load monitoring (NILM) is an elegant solution for monitoring energy consumption. NILM was getting popular since the advance of machine learning and deep learning technique. For the past years, there have been some deep learning techniques proposed for NILM. The results have shown that the performance of deep learning models can outperform the prior state of the art of NILM models such as Factorial Hidden Markov Model. A NILM model needs to identify distinctive power patterns of certain appliances in order to monitor the power consumptions. Statistical features (SFs) such as power difference and difference of variant power can be utilized to help the network learn better. As there is no single perfect model that can perfectly fit for everything, based on empirical research, we find that particular SF can be useful at certain type of load. This paper proposes a parallel bidirectional long short-term memory model with SFs to improve learning capability of the network. The proposed method is tested along with some most recent deep learning models on NILM such as DCNN, GLU-Res, BLSTM, and AE. The proposed method can successfully outperform those methods and shows consistent results.

This concluding chapter, while appreciating the disaggregating approach and emphasis on the local in many of the contributions to this volume, shifts the attention back to the broader national and especially international forces that determined the course of the Holocaust. Exclusive focus on the local and comparative microhistories, or even the internal features of a country or society, misses out on the important external factors that determined the causes and impact of the destruction of European Jewry. When democracy died in a regional hegemonic power, such as Germany in the 1930s, this paved the way for the horrors to follow well beyond its borders through force and emulation. Were democracy to die in a global hegemon such as the United States—an ominous possibility after 2016—“the reverberations would be massive.” Ultimately, the Holocaust offers an important warning: the local process of political violence and dehumanization highlighted in this book occurred when democracy died in a powerful nation.