High Performance Computing (HPC) and Cloud Computing datacenters are extensively used to steer and solve complex problems in science, engineering, and business, such as calculating correlations and making predictions. Already in a single datacenter server, there are thousands of hardware and software metrics – Key Performance Indicators (KPIs) – that individually and aggregated can give insight in the performance, robustness, and efficiency of the datacenter and the provisioned applications. At the datacenter level, the number of KPIs is even higher. The fast growing interest on datacenter management from both public and industry together with the rapid expansion in scale and complexity of datacenter resources and the services being provided on them have made monitoring, profiling, controlling, and provisioning compute resources dynamically at runtime into a challenging and complex task. Commonly, correlations of application KPIs, like response time and throughput, with resource capacities show that runtime systems (e.g., containers or virtual machines) that are used to provision these applications do not utilize available resources efficiently. This reduces datacenter efficiency, which in term results in higher operational costs and longer waiting times for results.

The goal of this thesis is to develop tools and autonomic techniques for improving datacenter operations, management and utilization, while improving and/or minimizing impacts on applications performance. To this end, we make use of application resource descriptors to create a library that dynamically adjusts the amount of resources used, enabling elasticity for scientific workflows in HPC datacenters. For mission critical applications, high availability is of great concern since these services must be kept running even in the event of system failures. By modeling and correlating specific resource counters, like CPU, memory and network utilization, with the number of runtime synchronizations, we present adaptive mechanisms to dynamically select which fault tolerant mechanism to use. Likewise, for scientific applications we propose a hybrid extensible architecture for dual-level scheduling of data intensive jobs in HPC infrastructures, allowing operational simplification, on-boarding of new types of applications and achieving greater job throughput with higher overall datacenter efficiency.

Over the last decade, new applications such as data intensive workflows have hit an inflection point in wide spread use and influenced the compute paradigm of most scientific and industrial endeavours. Data intensive workflows are highly dynamic and adaptable to resource changes, system faults, and by also allowing approximate solutions into their models. On the one hand, these dynamic characteristics require processing power and capabilities originated in cloud computing environments, and are not well supported by large High Performance Computing (HPC) infrastructures. On the other hand, cloud computing datacenters favor low latency over throughput, deeply contrasting with HPC, which enforces a centralized environment and prioritizes total computation accomplished over-time, ignoring latency entirely. Although data handling needs are predicted to increase by as much as a thousand times over the next decade, future datacenters processing power will not increase as much.

To tackle these long-term developments, this thesis proposes autonomic methods combined with novel scheduling strategies to optimize datacenter utilization while guaranteeing user defined constraints and seamlessly supporting a wide range of applications under various real operational scenarios. Leveraging upon data intensive characteristics, a library is developed to dynamically adjust the amount of resources used throughout the lifespan of a workflow, enabling elasticity for such applications in HPC datacenters. For mission critical environments where services must run even in the event of system failures, we define an adaptive controller to dynamically select the best method to perform runtime state synchronizations. We develop different hybrid extensible architectures and reinforcement learning scheduling algorithms that smoothly enable dynamic applications into HPC environments. An overall theme in this thesis is extensive experimentation in real datacenters environments. Our results show improvements in datacenter utilization and performance, achieving higher overall efficiency. Our methods also simplify operations and allow the onboarding of novel types of applications previously not supported.

Dataintensiva workflows är en ny klass av applikationer som blivit alltmer vanliga under senaste årtiondet och har stor påverkan på hur beräkningar utförs inom flertalet forskningsområden och i industrin. Dessa dataintensiva workflows kan dynamiskt anpassa sig till ändringar i resursallokering, systemfel och kan ibland även approximera lösningar vid resursbrist. De kräver hög beräkningskraft och därtill funktionalitet som endast återfinns i datormoln och de passar därmed dåligt i dagens högpresterande datorsystem (HPC-system). Datacenter i molnet prioriterar att snabbt starta nyinkomna applikationer, vilket drastiskt skiljer sig från HPC-miljöer där hög genomströmning över tid är det främsta målet. Trots att behovet av datahantering uppskattas öka mer än tusenfallt under kommande årtioende kommer framtidens datacenter inte att ha motsvarande utveckling av beräkningskapacitet.

Denna avhandling möter dessa utmaningar genom en kombination av autonoma system och nya strategier för schedulering för att optimera utnyttjandegraden i datacenter. Detta sker utan att göra avkall på användares prestandakrav och därtill med målet att stödja ett brett spektrum av applikationer och scenarios. Ett bibliotek utvecklas för att dynamiskt anpassa resursallokering för workflows under körning, vilket innebär att även HPC-system kan stödja elastiska applikationer som tidigare bara kunde exekveras i datormoln. För miljöer med höga krav på tillgänglighet defineras en regulator för att dynamiskt anpassa hur applikationer synkroniserar tillstånd, för mer resurseffektiv aktiv replikering. Avhandlingen utvecklar även flera resurshanteringssystem baserat på schedulering med förstärkningsinlärning i syftet att förbättra stödet för dynamiska applikationer i HPC-system. Ett övergripande tema i avhandlingen är omfattande utvärderingar av de framtagna metoderna och systemen genom storskaliga experiment i verkliga datacenter. Resultaten visar förbättringar överlag av resursutnyttjande och prestanda i datacenter. De utvecklade systemen förenklar även drift och möjliggör nya typer av applikationer som tidigare ej kunnat exekveras i HPC-miljöer.