Πρόσφατα η ιντερνετική κοινότητα έστρεψε την προσοχή και το ενδιαφέρον της στα peer-to-peer συστήματα, όπου οι χρήστες προσφέρουν τους πόρους τους (αποθηκευτικό χώρο, υπολογιστικό χρόνο) και περιεχόμενο (αρχεία) στην διάθεση της κοινότητας. Οι χρήστες θεωρούνται ίσοι μεταξύ τους και ο καθένας συμμετέχει με διπλό ρόλο, τόσο σαν πελάτης, όσο και σαν εξυπηρετητής. Με αυτό τον τρόπο δημιουργούνται κοινότητες με αποκεντρωμένο έλεγχο, ευελιξία, σταθερότητα, κλιμάκωση, ανωνυμία κι αντοχή στην λογοκρισία. Ενώ όμως οι δυνατότητες αυτών των συστημάτων είναι ποικίλες, την μεγαλύτερη αποδοχή έχουν γνωρίσει τα συστήματα ανταλλαγής αρχείων όπως το Napster, Kazaa, Gnutella, eDonkey, BitTorrent. Το ενδιαφέρον των ερευνητών δεν έχει περιοριστεί μόνο στην ανταλλαγή δεδομένων μιας και η ανοιχτή φύση των peer-to-peer συστημάτων προσφέρει πολύ περισσότερες προκλήσεις. Ένα ενδιαφέρον πεδίο έρευνας είναι αυτό των τεχνικών εξισορρόπησης φόρτου. Η έρευνα που διεξάγεται μπορεί να χωριστεί σε δύο κατηγορίες. Στην μια κατηγορία μπορούμε να εντάξουμε τεχνικές για την καλύτερη κατανομή των αντικειμένων στον χώρο ονομάτων για βελτιστοποιήσεις στην απόδοση της δρομολόγησης και αναζήτησης [Pastry, Tapestry, Chord]. Στην δεύτερη μπορούμε να εντάξουμε τεχνικές για την κατανομή αντιγράφων των αντικειμένων στους κόμβους του δικτύου, για βελτιστοποίηση του ρυθμού εξυπηρέτησης των χρηστών και της ποιότητας υπηρεσίας που τους προσφέρει το σύστημα, η οποία μπορεί άμεσα να συσχετιστεί με την διαθεσιμότητα. Ενώ η δεύτερη κατηγορία μπορούμε να πούμε ότι φαίνεται να είναι πιο ενδιαφέρουσα από την σκοπιά του τελικού χρήστη, η έρευνα στον τομέα αυτό δεν φαίνεται να λαμβάνει υπόψη της έναν ρεαλιστικό υπολογισμό του κόστους εφαρμογής των προτεινομένων τεχνικών. Έτσι κάποιες εργασίες δεν υπολογίζουν καθόλου το κόστος δημιουργίας αντιγράφων, ενώ κάποιες άλλες το θεωρούν σταθερό και ανεξάρτητο από το μέγεθος των αντικειμένων και τη σύνδεση των δυο κόμβων μεταξύ των οποίων γίνεται η επικοινωνία. Κάποιες λιγότερο ή περισσότερο προχωρούν λίγο παραπέρα και ορίζουν το κόστος να είναι ανάλογο του απαιτούμενου αποθηκευτικού χώρου. Η διαφορά με την παρούσα εργασία είναι ότι δεν εμπεριέχουν την έννοια της εξισορρόπησης του φόρτου των κόμβων μεταξύ τους. Σε αυτή την εργασία προσπαθήσαμε να καθορίσουμε ένα σύνολο μετρικών απόδοσης μέσα από ένα πλαίσιο εργασίας για έναν όσο το δυνατόν περισσότερο ρεαλιστικό τρόπο υπολογισμού τους. Για να το πετύχουμε αυτό, καταρχήν σχεδιάσαμε ένα p2p σύστημα διαμοίρασης αρχείων με αρχιτεκτονική που βασίζεται στην οργάνωση των κόμβων σε ομάδες, ενώ στη συνέχεια ορίζοντας την έννοια του φόρτου υλοποιήσαμε τεχνικές για την εξισορρόπησή του. Για την αξιολόγηση των τεχνικών, ορίστηκε ένα σύνολο μετρικών οι οποίες καταγράφουν την απόδοση του συστήματος τόσο από την οπτική γωνία του συστήματος (επιθυμητή η δίκαιη κατανομή του φόρτου και η βέλτιστη χρήση των πόρων όπως κυρίως η διαθέσιμη χωρητικότητα των συνδέσεων των κόμβων μεταξύ τους), όσο κι από την οπτική γωνία του χρήστη (καλύτερη «ποιότητα υπηρεσίας» με το ελάχιστο δυνατό κόστος). Η πειραματική αξιολόγηση των τεχνικών έγινε μέσα σε ένα περιβάλλον προσομοίωσης, το οποίο υλοποιήθηκε από μηδενική βάση, έπειτα από μια μελέτη παρόμοιων συστημάτων. Τα κύρια χαρακτηριστικά του περιβάλλοντος αυτού είναι α) η επεκτασιμότητα κι ευχρηστία του, β) απλή και ρεαλιστική βάση προσομοίωσης, γ) ύπαρξη πληθώρας παραμέτρων της προσομοίωσης που δίνονται σαν είσοδος από τον χρήστη και δ) δυνατότητα προσομοίωσης μεγάλου μεγέθους συστημάτων. / Recently the internet community has focused its interest on peer-to-peer systems, where users contribute their resources (storage space, computation time) and content (documents, files) to the community. The users are considered equivalent, each one participating with a dual role, both as a client and server. The communities formed under this simple peer-to-peer paradigm are characterized by the decentralized control, robustness, stability, scaling, anonymity and resistance to censorship. While there are various potential application domains of peer-to-peer systems, depending on the type of shared resources, the file sharing systems, such as Napster, Kazaa, Gnutella, eDonkey, BitTorrent, has known the greater acceptance. The “open nature” of peer-to-peer systems offers a wider area of research interest and much more challenges than just content sharing; interesting research domains include infrastructure, collaboration, searching, routing, load balancing, security etc Load balancing is a very interesting domain on such systems. The carried out research in this domain may be categorized in two categories. In the first, one can include techniques for better item distribution in the name space so as improvements in routing and searching can be accomplished [ref2 Pastry, Tapestry, Chord]. In the second, one can include techniques for items’ replicas placement to the network nodes, for improving the throughput and the Quality of Service provided to the users. The QoS can be straightforward related to the availability While the second category seems to be more interesting from the user’s perspective, the research in this domain does not seem to take into account a realistic cost evaluation of the proposed techniques. Some research studies just ignore it, while some others consider it constant and irrelative to the objects’ size and the connection between the two nodes where the object transfer occurs. Some others (less) (or more) get a little further and define the cost to be proportional to the needed storage capacity. The difference with our study is that the previous studies do not comprise the notion of load balancing among users as well as evaluate the cost under different assumptions. With our work we try to define a set of performance metrics through a framework based on a measurement as realistic as possible. To accomplish this, at first we designed a cluster based file sharing p2p system, then we defined the notion of load and finally implemented load balancing techniques. To evaluate these techniques we defined a set of metrics that record the system’s performance both from the system’s perspective (desirable the fair load distribution and the optimum use of resources like the available bandwidth of nodes’ connections) and the user’s (better “quality of service” with the least cost). For the experimental evaluation of these techniques we developed from scratch a simulation environment, after we studied similar systems. The main characteristics of this simulator are a) extensibility and usability, b) simple and realistic simulation base, c) availability of plenty simulation parameters given as input from the user d) scalability to simulate large scale systems.

Random graph processes are most often used to investigate theoretical questions about random graphs. A randomised algorithm can be defined specifically for the purpose of finding some structure in a graph, such as a matching, a colouring or a particular kind of sub graph. Properties of the related random graph process then suggest properties, or bounds on properties, of the structure. In this thesis, we use a random graph process to analyse a particular load balancing algorithm from theoretical computer science. By doing so, we demonstrate that random graph processes may also be used to analyse other algorithms and systems of a random nature, from areas such as computer science, telecommunications and other areas of engineering and mathematics. Moreover, this approach can lead to theoretical results on the performance of algorithms that are difficult to obtain by other methods. In the course of our analysis we are also led to some results of the first kind, relating to the structure of the random graph. / The particular algorithm that we analyse is a randomised algorithm for an off-line load balancing problem with two choices. The load balancing algorithm, in an initial stage, mirrors an algorithm which finds the k-core of a graph. This latter algorithm and the related random graph process have been previously analysed by Pittel, Spencer and Wormald, using a differential equation method, to determine the thresholds for the existence of a k-core in a random graph. We modify their approach by using a random pseudograph model due to Bollobas and Frieze, and Chvatal, in place of the uniform random graph. This makes the analysis somewhat simpler, and leads to a shortened derivation of the thresholds and other properties of k-cores.(For complete abstract open document)

This thesis explores the creation of a scaling, containerized, advertisement server that will be used by Gold Town Games AB to better integrate ads into their application(s). The server is built as a Docker image that will be used to create server instances on AWS Elastic Container Service for automatic scaling and server resource configuration. The server was created with the intention that GTG will have full control over what advertisements are shown in their application(s) and to seamlessly integrate sponsored logos onto jerseys or sports fields. This will not only serve as a source of income with advertisers paying for ad space, but it will also make the game elements more realistic as we have come to expect teams and stadiums to be sponsored and plastered with company logos. Another important part when displaying advertisement is to track statistics for the ads, since without a way to show advertisers that their ads are shown and that they are generating engagement it is very hard to sell the ad space. / Detta examensarbete utforskar skapandet av en skalbar, containerbaserad, reklamserver som kommer användas av Gold Town Games AB för att integrera reklam i deras applikation(er). Servern är byggd som en Docker-bild som används för att skapa instanser på AWS Elastic Container Service för automatisk skalning och serverresurshantering. Servern är utvecklad med tanken att GTG ska ha full kontroll över vilken reklam som visas i deras applikation(er) och för att kunna lägga till sponsrade loggor på matchtröjor och i arenor. Detta är inte bara en extra form av inkomst, då annonsörer betalar för reklamplatser, utan hjälper även till att få delar av spelen att kännas mer realistiska då vi är vana att lag och arenor är sponsrade och fulla av företagsloggor. En annan viktig del när man visar reklam är att kunna spara statistik för den, eftersom det skulle vara väldigt svårt att sälja reklamplatser utan att kunna visa att folk faktiskt ser reklamen.

L’émergence des applications de bases de données dans les domaines tels que le data warehousing,le data mining et l’aide à la décision qui font généralement appel à de très grands volumes de donnéesrend la parallélisation des algorithmes des jointures nécessaire pour avoir un temps de réponse acceptable.Une accélération linéaire est l’objectif principal des algorithmes parallèles, cependant dans les applicationsréelles, elle est difficilement atteignable : ceci est dû généralement d’une part aux coûts de communicationsinhérents aux systèmes multi-processeurs et d’autre part au déséquilibre des charges des différents processeurs.En plus, dans un environnement hétérogène multi-utilisateur, la charge des différents processeurspeut varier de manière dynamique et imprévisible.Dans le cadre de cette thèse, nous nous intéressons au traitement de la jointure et de la multi-jointure surles architectures distribuées hétérogènes, les grilles de calcul et les systèmes de fichiers distribués. Nousavons proposé une variété d’algorithmes, basés sur l’utilisation des histogrammes distribués, pour traiterde manière efficace le déséquilibre des données, tout en garantissant un équilibrage presque parfait dela charge des différents processeurs même dans un environnement hétérogène et multi-utilisateur. Cesalgorithmes sont basés sur une approche dynamique de redistribution des données permettant de réduire lescoûts de communication à un minimum tout en traitant de manière très efficace le problème de déséquilibredes valeurs de l’attribut de jointure.L’analyse de complexité de nos algorithmes et les résultats expérimentaux obtenus montrent que cesalgorithmes possèdent une accélération presque linéaire. / The appeal of parallel processing becomes very strong in applications which require ever higher performanceand particularly in applications such as : data-warehousing, decision support, On-Line Analytical Processing(OLAP) and more generally DBMS. A linear speed-up is the main objective of parallel algorithms. However,in real applications, it’s not obvious to reach this objective due to the high communication cost in parallel anddistributed systems and to the possible skew in the charge of different processors. In addition, on heterogeneousmulti-user architectures, the load of each processor may highly vary in a dynamic and unpredictableway.In this thesis, we are interested in treating the join and multi-join queries on distributed multi-user heteregeneoussystems, grid systems and distributed file systems. We have proposed several algorithms based onusing distributed histograms. These algorithms are based on a dynamic data distribution and task allocationwhich makes them insensitive to data skew and ensure perfect balancing properties during all stages of joincomputation even on heteregeneous multi-user environment. The complexity analysis of our algorithms andthe experimental results show that they have a near-linear speedup.

Weather forecasting models are computationally intensive applications and traditionally they are executed in parallel machines. However, some issues prevent these models from fully exploiting the available computing power. One of such issues is load imbalance, i.e., the uneven distribution of load across the processors of the parallel machine. Since weather models are typically synchronous applications, that is, all tasks synchronize at every time-step, the execution time is determined by the slowest task. The causes of such imbalance are either static (e.g. topography) or dynamic (e.g. shortwave radiation, moving thunderstorms). Various techniques, often embedded in the application’s source code, have been used to address both sources. However, these techniques are inflexible and hard to use in legacy codes. In this thesis, we explore the concept of processor virtualization for dynamically balancing the load in weather models. This means that the domain is over-decomposed in more tasks than the available processors. Assuming that many tasks can be safely executed in a single processor, each processor is put in charge of a set of tasks. In addition, the system can migrate some of them from overloaded processors to underloaded ones when it detects load imbalance. This approach has the advantage of decoupling the application from the load balancing strategy. Our objective is to show that processor virtualization can be applied to weather models as long as an appropriate strategy for migrations is used. Our proposal takes into account the communication pattern of the application in addition to the load of each processor. In this text, we present the techniques used to minimize the amount of change needed in order to apply processor virtualization to a real-world application. Furthermore, we analyze the effects caused by the frequency at which the load balancer is invoked and a threshold that activates rebalancing. We propose an automatic strategy to find an optimal threshold to trigger load balancing. These strategies are centralized and work well for moderately large machines. For larger machines, we present a fully distributed algorithm and analyze its performance. As a study case, we demonstrate the effectiveness of our approach for dynamically balancing the load in Brams, a mesoscale weather forecasting model based on MPI parallelization. We choose this model because it presents a considerable load imbalance caused by localized thunderstorms. In addition, we analyze how other effects of processor virtualization can improve performance.

Dans le contexte du calcul scientique, l'équilibrage de la charge est un problème crucial qui conditionne la performance des simulations numériques parallèles. L'objectif est de répartir la charge de travail entre un nombre de processeurs donné, afin de minimiser le temps global d'exécution. Une stratégie populaire pour résoudre ce problème consiste à modéliser la simulation à l'aide d'un graphe et à appliquer des algorithmes de partitionnement. En outre, les simulations numériques tendent à se complexifier, notamment en mixant plusieurs codes représentant des physiques différentes ou des échelles différentes. On parle alors de couplage de codes multi-physiques ou multi-échelles. Dans ce contexte, le problème de l'équilibrage de charge devient également plus difficile, car il ne s'agit plus d'équilibrer chacun des codes séparément, mais l'ensemble de ces codes pris dans leur globalité. Dans ce travail, on propose de resoudre ce problème en utilisant le modèle de partitionnement à sommets fixes qui pourrait représenter efficacement les contraintes supplémentaires imposées par les codes couplés (co-partitionnement). Nous avons donc développé un algorithme direct de partitionnement de graphe qui gère des sommets fixes. L'algorithme a été implémenté dans le partitionneur Scotch et une série d'expériences ont été menées sur la collection des graphes DIMACS. Ensuite nous avons proposé trois algorithmes de co-partitionnement qui respectent les contraintes issues des codes couplés respectifs. Nous avons egalement validé nos algorithmes par une étude expérimentale en comparant nos méthodes aux strategies actuelles sur des cas artificiels ainsi que sur des codes réels couplés. / Load balancing is an important step conditioning the performance of parallel applications. The goal is to distribute roughly equal amounts of computational load across a number of processors, while minimising interprocessor communication. A common approach to model the problem is based on graph structures and graph partitioning algorithms. Moreover, new challenges involve the simulation of more complex physical phenomena, where different parts of the computational domain exhibit different physical behavior. Such simulations follow the paradigm of multi-physics or multi-scale modeling approaches. Combining such different models in massively parallel computations is still a challenge to reach high performance. Additionally, traditional load balancing algorithms are often inadequate, and more sophisticated solutions should be explored. In this thesis, we propose new graph partitioning algorithms that balance the load of such simulations, refered to as co-partitioning. We formulate this problem with the use of graph partitioning with initially fixed vertices which we believe represents efficiently the additional constraints of coupled simulations. We have therefore developed a direct algorithm for graph partitioning that manages successfully problems with fixed vertices. The algorithm is implemented inside Scotch partitioner and a series of experiments were carried out on the DIMACS graph collection. Moreover we proposed three copartitioning algorithms that respect the constraints of the respective coupled codes. We finally validated our algorithms by an experimental study comparing our methods with current strategies on artificial cases and on real-life coupled simulations.

Weather forecasting models are computationally intensive applications and traditionally they are executed in parallel machines. However, some issues prevent these models from fully exploiting the available computing power. One of such issues is load imbalance, i.e., the uneven distribution of load across the processors of the parallel machine. Since weather models are typically synchronous applications, that is, all tasks synchronize at every time-step, the execution time is determined by the slowest task. The causes of such imbalance are either static (e.g. topography) or dynamic (e.g. shortwave radiation, moving thunderstorms). Various techniques, often embedded in the application’s source code, have been used to address both sources. However, these techniques are inflexible and hard to use in legacy codes. In this thesis, we explore the concept of processor virtualization for dynamically balancing the load in weather models. This means that the domain is over-decomposed in more tasks than the available processors. Assuming that many tasks can be safely executed in a single processor, each processor is put in charge of a set of tasks. In addition, the system can migrate some of them from overloaded processors to underloaded ones when it detects load imbalance. This approach has the advantage of decoupling the application from the load balancing strategy. Our objective is to show that processor virtualization can be applied to weather models as long as an appropriate strategy for migrations is used. Our proposal takes into account the communication pattern of the application in addition to the load of each processor. In this text, we present the techniques used to minimize the amount of change needed in order to apply processor virtualization to a real-world application. Furthermore, we analyze the effects caused by the frequency at which the load balancer is invoked and a threshold that activates rebalancing. We propose an automatic strategy to find an optimal threshold to trigger load balancing. These strategies are centralized and work well for moderately large machines. For larger machines, we present a fully distributed algorithm and analyze its performance. As a study case, we demonstrate the effectiveness of our approach for dynamically balancing the load in Brams, a mesoscale weather forecasting model based on MPI parallelization. We choose this model because it presents a considerable load imbalance caused by localized thunderstorms. In addition, we analyze how other effects of processor virtualization can improve performance.

Em um sistema localmente distribuído, quando diversas tarefas são aleatoriamente submetidas aos computadores de uma determinada rede, onde alguns ficam sobrecarregados e outros permaneçam ociosos, algoritmos de balanceamento de carga podem ser utilizados para homogeneizar e otimizar a alocação de recursos e, conseqüentemente, aumentar o desempenho computacional. Com a evolução dos sistemas distribuídos, houve a necessidade de aprimoramento desses algoritmos, a fim de suportarem sistemas distribuídos altamente escaláveis e possibilitarem o gerenciamento de ambientes heterogêneos. Visando suprir essas necessidades, recentemente foi proposto um algoritmo de balanceamento de carga denominado Tree Load Balancing Algorithm (TLBA). Para avaliar esse algoritmo foram desenvolvidos um simulador e um protótipo. Com as simulações foram comprovados seus benefícios para ambientes distribuídos heterogêneos altamente escaláveis e com o protótipo esses resultados foram validados. Este trabalho apresenta a implementação de um novo simulador, que possibilitou um estudo mais abrangente dos parâmetros de balanceamento de carga do TLBA, com base em amostras estatisticamente geradas e uma maior fidelidade em suas políticas de balanceamento. Suas novas características são simulação em tempo real, visualização da árvore lógica de acordo com as capacidades computacionais relativas, resultados em tabelas, gráficos gerados pelo simulador e estudos aplicados aos diferentes tipos de escalonamento e sistemas / In locally distributed systems, when plenty of tasks are randomically submitted to network computers where certain computers may be heavily loaded, while others lightly loaded, load balancing algorithms may be used on to homogenize and optimize the resources allocation, and hence improve the computer performance. Later on, the distributed systems evolution made necessary the upgrading of these algorithms in order to support the highly scalable distributed systems and handle with the management of the heterogeneous environments. To supply these needs, a new load balancing algorithm called Tree Load Balancing Algorithm (TLBA) has been recently proposed. The evaluation of this algorithm has been done by a simulator and a prototype which were developed for this purpose. The simulations have proved its benefits in highly scalable heterogeneous distributed environments, and with the prototype the results were validated. This work presents the implementation of a new simulator which allows detailed studies of the TLBA parameters, based on the statistically generated samples, and a higher fidelity on the implementation of its load balancing policies. Its new features are the real time simulations, visualization of the logical tree according to computers’ relative capacities, results in tables, graphics generated by the simulator and studies applied to the different types of scheduling and systems

Weather forecasting models are computationally intensive applications and traditionally they are executed in parallel machines. However, some issues prevent these models from fully exploiting the available computing power. One of such issues is load imbalance, i.e., the uneven distribution of load across the processors of the parallel machine. Since weather models are typically synchronous applications, that is, all tasks synchronize at every time-step, the execution time is determined by the slowest task. The causes of such imbalance are either static (e.g. topography) or dynamic (e.g. shortwave radiation, moving thunderstorms). Various techniques, often embedded in the application’s source code, have been used to address both sources. However, these techniques are inflexible and hard to use in legacy codes. In this thesis, we explore the concept of processor virtualization for dynamically balancing the load in weather models. This means that the domain is over-decomposed in more tasks than the available processors. Assuming that many tasks can be safely executed in a single processor, each processor is put in charge of a set of tasks. In addition, the system can migrate some of them from overloaded processors to underloaded ones when it detects load imbalance. This approach has the advantage of decoupling the application from the load balancing strategy. Our objective is to show that processor virtualization can be applied to weather models as long as an appropriate strategy for migrations is used. Our proposal takes into account the communication pattern of the application in addition to the load of each processor. In this text, we present the techniques used to minimize the amount of change needed in order to apply processor virtualization to a real-world application. Furthermore, we analyze the effects caused by the frequency at which the load balancer is invoked and a threshold that activates rebalancing. We propose an automatic strategy to find an optimal threshold to trigger load balancing. These strategies are centralized and work well for moderately large machines. For larger machines, we present a fully distributed algorithm and analyze its performance. As a study case, we demonstrate the effectiveness of our approach for dynamically balancing the load in Brams, a mesoscale weather forecasting model based on MPI parallelization. We choose this model because it presents a considerable load imbalance caused by localized thunderstorms. In addition, we analyze how other effects of processor virtualization can improve performance.

Este trabalho apresenta a implementação e a utilização da migração de processos SPMD (Single Program Multiple Data), a qual realiza somente a transferência dos dados, que estão sendo manipulados pelo processo, para realizar a migração. Seu principal objetivo foi o estudo do impacto do balanceamento de carga no desempenho de aplicações, desenvolvidas utilizando o modelo de programação SPMD. Depois de realizados testes com aplicações SPMD reais, em sistemas computacionais distribuídos utilizando a migração de processos SPMD, foi possível verificar que ganhos de desempenho podem ser alcançados, tanto na migração de processos quanto no tempo de execução de aplicações paralelas SPMD. / This research presents the implementation and use of the SPMD (Single Program Multiple Data) process migration, which only does the transference of the data that are being used by the process, to perform the process migration. Its main objective was the study of the load balancing impact in the performance of applications developed using the SPMD programming model. After performing the tests with real SPMD applications, in distributed computational systems using the SPMD process migration, it was achieved good performance gains, both in the process migration and in the time execution of applications SPMD parallel applications.