Różnice

Różnice między wybraną wersją a wersją aktualną.

Odnośnik do tego porównania

2019:figiela:start [2019/06/17 10:21]
Kamil Figiela
2019:figiela:start [2019/06/17 10:34] (aktualna)
Kamil Figiela
Linia 23: Linia 23:
  
  
-**Streszczenie**+**Abstract**
  
 +In this thesis, we address a problem of execution scientific applications in the cloud environment. Optimal provisioning of resources and scheduling of tasks enable to achieve better turnaround times and keep infrastructure costs down.
  
 +First, we propose a state of the art Mixed Integer Linear Programming (MILP) model for scheduling large scale scientific workflows on Infrastructure-as-a-Service platforms. Our heuristic minimizes the cost of workflow execution under deadline constraint. Resources are provisioned across multiple providers for a single run of the workflow. We account for the data transfer cost of cross-provider communication and consider infrastructure characteristics that are frequently omitted such as per-hour billing or resource quotas. The algorithm is evaluated on synthetic workflows representing real scientific applications, such as Montage or CyberShake. We show that MILP may be efficiently used for scheduling scientific workloads. By solving the optimization problem for multiple deadlines, we produce trade-off plots, that show how the cost depends on the deadline. Such plots are a step towards a scientific cloud workflow calculator, supporting resource management decisions for both end-users and workflow-as-a-service providers.
 +
 +Later, we demonstrate that computation structure analysis and smart aggregation of tasks enables for efficient scheduling and execution of multi-frontal solver, an application typically considered as HPC that can be represented as a workflow. Specifically, we show how multi-frontal solver can be ported to be efficiently executed in loosely coupled cloud architecture. We execute the workflow using Hyperflow workflow engine and perform experiments in the real production environment. We compare results with deployment on an HPC cluster. The conclusion is that while the clouds are not as efficient in terms of raw performance as HPC systems, they can provide a better turnaround time, reducing the "time to science" which is the ultimate metric of efficiency important for the end user.
 +
 +Finally, we present a novel approach to performance evaluation of serverless infrastructure. We implement a performance evaluation framework and gather experimental data over a long period of time. We present a study on performance, heterogeneity, and behavior of runtime environment on Function-as-a-Service platforms. We determine differences in computational performance characteristics on leading commercial serverless platforms. We conclude that serverless infrastructures are economically viable for some cases of scientific workloads.
 +
 +
 +**Streszczenie**
  
 Rozprawa jest poświęcona zagadnieniu wykonania aplikacji naukowych w środowiskach chmurowych. Optymalny dobór infrastruktury i szeregowanie zadań pozwala na szybsze osiąganie rezultatów przy zachowaniu niskiego kosztu infrastruktury. Rozprawa jest poświęcona zagadnieniu wykonania aplikacji naukowych w środowiskach chmurowych. Optymalny dobór infrastruktury i szeregowanie zadań pozwala na szybsze osiąganie rezultatów przy zachowaniu niskiego kosztu infrastruktury.
2019/figiela/start.txt · ostatnio zmienione: 2019/06/17 10:34 przez Kamil Figiela