Du fait de la croissance de la demande en ressources de calcul, les architectures de centres de données gagnent en taille et complexité.Dès lors, cette thèse prend du recul par rapport aux architectures réseaux traditionnelles, et montre que fournir des primitives génériques directement à la couche réseau permet d'améliorer l'utilisation des ressources, et de diminuer le trafic réseau et le surcoût administratif.Deux architectures réseaux récentes, Segment Routing (SR) et Bit-Indexed Explicit Replication (BIER), sont utilisées pour construire et analyser des protocoles de couche réseau, afin de fournir trois primitives: (1) mobilité des tâches, (2) distribution fiable de contenu, et (3) équilibre de charge.Premièrement, pour la mobilité des tâches, SR est utilisé pour fournir un service de migration de machine virtuelles sans perte.Cela ouvre l'opportunité d'étudier comment orchestrer le placement et la migration de tâches afin de (i) maximiser le débit inter-tâches, tout en (ii) maximisant le nombre de nouvelles tâches placées, mais (iii) minimisant le nombre de tâches migrées.Deuxièmement, pour la distribution fiable de contenu, BIER est utilisé pour fournir un protocole de multicast fiable, dans lequel les retransmissions de paquets perdus sont ciblés vers l'ensemble précis de destinations n'ayant pas reçu ce packet : ainsi, le surcoût de trafic est minimisé.Pour diminuer la charge sur la source, cette approche est étendue en rendant possible des retransmissions par des pairs locaux, utilisant SR afin de trouver un pair capable de retransmettre.Troisièmement, pour l'équilibre de charge, SR est utilisé pour distribuer des requêtes à travers plusieurs applications candidates, chacune prenant une décision locale pour accepter ou non ces requêtes, fournissant ainsi une meilleure équité de répartition comparé aux approches centralisées.La faisabilité d'une implémentation matérielle de cette approche est étudiée, et une solution (utilisant des canaux cachés pour transporter de façon invisible de l'information vers l'équilibreur) est implémentée pour une carte réseau programmable de dernière génération.Finalement, la possibilité de fournir de l'équilibrage automatique comme service réseau est étudiée : en faisant passer (avec SR) des requêtes à travers une chaîne fixée d'applications, l'équilibrage est initié par la dernière instance, selon son état local. / With the development of demand for computing resources, data center architectures are growing both in scale and in complexity.In this context, this thesis takes a step back as compared to traditional network approaches, and shows that providing generic primitives directly within the network layer is a great way to improve efficiency of resource usage, and decrease network traffic and management overhead.Using recently-introduced network architectures, Segment Routing (SR) and Bit-Indexed Explicit Replication (BIER), network layer protocols are designed and analyzed to provide three high-level functions: (1) task mobility, (2) reliable content distribution and (3) load-balancing.First, task mobility is achieved by using SR to provide a zero-loss virtual machine migration service.This then opens the opportunity for studying how to orchestrate task placement and migration while aiming at (i) maximizing the inter-task throughput, while (ii) maximizing the number of newly-placed tasks, but (iii) minimizing the number of tasks to be migrated.Second, reliable content distribution is achieved by using BIER to provide a reliable multicast protocol, in which retransmissions of lost packets are targeted towards the precise set of destinations having missed that packet, thus incurring a minimal traffic overhead.To decrease the load on the source link, this is then extended to enable retransmissions by local peers from the same group, with SR as a helper to find a suitable retransmission candidate.Third, load-balancing is achieved by way of using SR to distribute queries through several application candidates, each of which taking local decisions as to whether to accept those, thus achieving better fairness as compared to centralized approaches.The feasibility of hardware implementation of this approach is investigated, and a solution using covert channels to transparently convey information to the load-balancer is implemented for a state-of-the-art programmable network card.Finally, the possibility of providing autoscaling as a network service is investigated: by letting queries go through a fixed chain of applications using SR, autoscaling is triggered by the last instance, depending on its local state.

Non-equilibrium surface growth for competitive growth models in (1+1) dimensions, particularly mixing random deposition (RD) with correlated growth process which occur with probability $p$ are studied. The composite mixtures are found to be in the universality class of the correlated growth process, and a nonuniversal exponent $\delta$ is identified in the scaling in $p$. The only effects of the RD admixture are dilations of the time and height scales which result in a slowdown of the dynamics of building up the correlations. The bulk morphology is taken into account and is reflected in the surface roughening, as well as the scaling behavior. It is found that the continuum equations and scaling laws for RD added, in particular, to Kardar-Parisi-Zhang (KPZ) processes are partly determined from the underlying bulk structures. Nonequilibrium surface growth analysis are also applied to a study of the static and dynamic load balancing for a conservative update algorithm for Parallel Discrete Event Simulations (PDES). This load balancing is governed by the KPZ equation. For uneven load distributions in conservative PDES simulations, the simulated (virtual) time horizon (VTH) per Processing Element (PE) and the imulated time horizon per volume element $N_{v}$ are used to study the PEs progress in terms of utilization. The width of these time horizons relates to the desynchronization of the system of processors, and is related to the memory requirements of the PEs. The utilization increases when the dynamic, rather than static, load balancing is performed.

No description available.

Edge computing, as an approach to leveraging computation capabilities located in different places, is widely deployed in the industry nowadays. With the development of edge computing, many big companies move from the traditional monolithic software architecture to the microservice design. To provide better performance of the applications which contain numerous loosely coupled modules that are deployed among multiple clusters, service routing among multiple clusters needs to be effective. However, most existing solutions are dedicated to static service routing and load balancing strategy, and thus the performance of the application cannot be effectively optimized when network condition changes.To address the problem mentioned above, we proposed a dynamic weighted round robin algorithm and implemented it on top of the cutting edge service mesh Istio. The solution is implemented as a Docker image called RoutingAgent, which is simple to deployed and managed. With the RoutingAgent running in the system, the weights of the target routing clusters will be dynamically changed based on the detected inter-cluster network latency. Consequently, the client-side request turnaround time will be decreased.The solution is evaluated in an emulated environment. Compared to the Istio without RoutingAgent, the experiment results show that the client-side latency can be effectively minimized by the proposed solution in the multicluster environment with dynamic network conditions. In addition to minimizing response time, emulation results demonstrate that loads of each cluster are well balanced. / Edge computing, som ett tillvägagångssätt för att utnyttja beräkningsfunktioner som finns på olika ställen, används i stor utsträckning i branschen nuförtiden. Med utvecklingen av kantdatabasen flyttar många stora företag från den traditionella monolitiska mjukvaruarkitekturen till mikroserviceteknik. För att ge bättre prestanda för de applikationer som innehåller många löst kopplade moduler som distribueras bland flera kluster, måste service routing bland flera kluster vara effektiva. De flesta befintliga lösningarna är dock dedikerade till statisk service-routing och belastningsbalanseringsstrategi, vilket gör att programmets prestanda inte effektivt kan optimeras när nätverksförhållandena ändras.För att ta itu med problemet som nämnts ovan föreslog vi en dynamisk viktad round robin-algoritm och implementerade den ovanpå den avancerade servicenätverket Istio. Lösningen implementeras som en Docker-bild som heter RoutingAgent, som är enkel att distribuera och hantera. Med agenten som körs i systemet ändras vikten av målruteringsklustret dynamiskt baserat på den upptäckta interklusternätets latens. Följaktligen kommer klientsidans begäran om omställningstid att minskas.Lösningen utvärderas i en emulerad miljö. Jämfört med Istio utan agent visar experimentresultaten att klientens latentitet effektivt kan minimeras av den föreslagna lösningen i multicluster-miljö med dynamiska nätverksförhållanden. Förutom att minimera responstid visar emuleringsresultat att belastningar i varje kluster är välbalanserade.

QUIC is an emerging connection-oriented transport layer protocol that aims to support low-latency and highly secure communication between users and cloud services. Cloud load balancers distribute incoming QUIC connections towards a pool of backend servers where applications run. The sheer size of today’s clouds requires to deploy large pools of load balancing instances across multi-core servers. Today, once a connection is handled by a load balancing instance, the handling of the connection cannot be delegated to another instance which is possibly running on an underutilized server or core. Recent work has shown that such imbalances lead to poor utilization of the server resources when implementing the load balancer component itself. Moving the handling of a connection from one load balancing instance to another one is key to achieve uniform distribution of the workload among the load balancing instances. This thesis discusses the system design, algorithmic principles, and evaluation of an approach to move the state of a QUIC connection from one QUIC instance to another one, a task that we call “connection migration”. Our preliminary investigation focuses on inter-CPU-core thread-based migration of QUIC connections within the same server machine with the goal of supporting different types of load balancers that dynamically spread the workload across the load balancing instances. We evaluated the performance of the proposed implementation in a real physical testbed in NSLab at KTH across different dimensions: the number of active connections, the size of the file transfers within each connection, and the number of load balancing instances. The results show that under ideal circumstances, the throughput of a server system built by an 8-core computer reaches 100MB per second and handles up to 500 client requests per second. Finally, this thesis analyzes the bottleneck of the system and the critical parts of the implementation that should be optimized for better performance. / QUIC är ett framväxande anslutningsorienterat protokoll för transportlager som syftar till att stödja låg latens och mycket säker kommunikation mellan användare och molntjänster. Molnslastbalansering distribuerar inkommande QUIC-anslutningar mot en pool av Backend-servrar där applikationer körs. Den stora storleken på dagens moln kräver att stora pooler av lastbalanseringsinstanser distribueras över flerkärniga servrar. Idag, när en anslutning hanteras av en lastbalanseringsinstanser, kan hanteringen av anslutningen inte delegeras till en annan instans som möjligen körs på en underutnyttjad server eller kärna. Senaste arbetet har visat att sådana obalanser leder till dåligt utnyttjande av serverresurserna vid implementering av själva lastbalanseringskomponenten. Att flytta hanteringen av en anslutning från en lastbalanseringsinstans till en annan är nyckeln för att uppnå enhetlig fördelning av arbetsbelastningen mellan lastbalanseringsinstanserna. Denna avhandling diskuterar systemdesign, algoritmiska principer och utvärdering av en metod för att flytta tillståndet för en QUIC-anslutning från en QUIC-instans till en annan, en uppgift som vi kallar “anslutningsmigrering”. Vår preliminära undersökning fokuserar på intel-CPU-core trådbaserad migration av QUIC-anslutningar inom samma servermaskin med målet att stödja olika typer av belastningsutjämnare som dynamiskt sprider arbetsbelastningen över lastbalanseringsinstanser. Vi utvärderade prestanda för den föreslagna implementeringen i en riktig fysisk testbädd i NSLab vid KTH över olika dimensioner: antalet aktiva anslutningar, storleken på filöverföringarna inom varje anslutning och antalet lastbalanseringsinstanser. Resultaten visar att under idealiska omständigheter når genomströmningen av ett serversystem som byggts av en 8-kärnig dator 100 MB per sekund och hanterar upp till 500 klientförfrågningar per sekund. Slutligen analyserar denna avhandling systemets flaskhals och de kritiska delarna av implementeringen som bör optimeras för bättre prestanda.

Cloud computing is increasingly attracting huge attention both in academic research and industry initiatives and has been widely used to solve advanced computation problem. As cloud datacentres continue to grow in scale and complexity, the risk of failure of Virtual Machines (VM) and hosts running several jobs and processing large amount of user request increases and consequently becomes even more difficult to predict potential failures within a datacentre. However, even though fault tolerance continues to be an issue of growing concern in cloud and HPC systems, mitigating the impact of failure and providing accurate predictions with enough lead time remains a difficult research problem. Traditional existing fault-tolerance strategies such as regular check-point/restart and replication are not adequate due to emerging complexities in the systems and do not scale well in the cloud due to resource sharing and distributed systems networks. In the thesis, a new reliable Fault Tolerance scheme using an intelligent optimal strategy is presented to ensure high system availability, reduced task completion time and efficient VM allocation process. Specifically, (i) A generic fault tolerance algorithm for cloud data centres and HPC systems in the cloud was developed. (ii) A verification process is developed to a fully dimensional VM specification during allocation in the presence of fault. In comparison to existing approaches, the results obtained shows an increase in success rate of the VMs, a reduction in response time of VM allocation and an improved overall performance. (iii) A failure prediction model is further developed, and the predictive capabilities of machine learning is explored by applying several algorithms to improve the accuracy of prediction. Experimental results indicate that the average prediction accuracy of the proposed model when predicting failure is about 90% accurate compared to existing algorithms, which implies that the approach can effectively predict potential system and application failures within the system.

No description available.

No description available.

No description available.

High performance computing (HPC) is routinely employed in diverse domains such as life sciences, and Geology, to simulate and understand the behavior of complex phenomena. Big data driven scientific simulations are resource intensive and require both computing and I/O capabilities at scale. There is a crucial need for revisiting the HPC I/O subsystem to better optimize for and manage the increased pressure on the underlying storage systems from big data processing. Extant HPC storage systems are designed and tuned for a specific set of applications targeting a range of workload characteristics, but they lack the flexibility in adapting to the ever-changing application behaviors. The complex nature of modern HPC storage systems along with the ever-changing application behaviors present unique opportunities and engineering challenges. In this dissertation, we design and develop a framework for optimizing HPC storage systems by making them application-attuned. We select three different kinds of HPC storage systems - in-memory data analytics frameworks, parallel file systems and object storage. We first analyze the HPC application I/O behavior by studying real-world I/O traces. Next we optimize parallelism for applications running in-memory, then we design data management techniques for HPC storage systems, and finally focus on low-level I/O load balance for improving the efficiency of modern HPC storage systems. / Doctor of Philosophy / Clusters of multiple computers connected through internet are often deployed in industry and laboratories for large scale data processing or computation that cannot be handled by standalone computers. In such a cluster, resources such as CPU, memory, disks are integrated to work together. With the increase in popularity of applications that read and write a tremendous amount of data, we need a large number of disks that can interact effectively in such clusters. This forms the part of high performance computing (HPC) storage systems. Such HPC storage systems are used by a diverse set of applications coming from organizations from a vast range of domains from earth sciences, financial services, telecommunication to life sciences. Therefore, the HPC storage system should be efficient to perform well for the different read and write (I/O) requirements from all the different sets of applications. But current HPC storage systems do not cater to the varied I/O requirements. To this end, this dissertation designs and develops a framework for HPC storage systems that is application-attuned and thus provides much improved performance than other state-of-the-art HPC storage systems without such optimizations.