Post-graduate theses
Current Record: 78 of 824
|
|
| Identifier |
000441080 |
| Title |
Elastic resource allocation for a structural design application |
| Alternative Title |
Ελαστική διανομή πόρων για εφαρμογή στατικής ανάλυσης κτιρίων |
|
Author
|
Γιορταμής, Εμμανουήλ
|
|
Thesis advisor
|
Πρατικάκης, Πολύβιος
|
|
Reviewer
|
Μπίλας, Άγγελος
Μαγκούτης, Κωνσταντίνος
|
| Abstract |
Cloud computing is the on-demand delivery of computing services such as servers,
storage, computing power, databases and networking, through the internet. Rather that
owning their own infrastructure, individuals or companies can rent access to highperformant computing resources from a cloud service provider. Cloud providers typically
offer pay-as-you-go pricing models which charge users only for the resources they use.
However, applications tend to have varying resource demands depending on both
incoming traffic rates and incoming workload types. Resource over-allocation leads to
wasted resources and thus, money, while under-allocation leads to Service-LevelObjective (SLO) violations. To this end, cloud computing platforms adopt horizontal and
vertical elasticity in order to timely scale the application's resources on demand.
Horizontal elasticity replicates the application's resources while vertical elasticity resizes
them. It came to our attention that both industry and scientific literature focus more on
horizontal elasticity than on vertical elasticity. Vertically elastic resource scaling is
essential for applications with workload-dependent and spiky resource demands,
however.
In this thesis we present a vertically elastic resource allocator for fine-grained CPU-time
allocation. Our proposed algorithm targets applications with job dependent
parallelization spikes and accounts for variable traffic rates. Our example application is a
Greek commercial structural design application used by civil engineers, named RAF. Its
back-end, RAF::Solver, computes a building's static analyses by solving linear algebra
equations and factorizing matrices using parallel Cholesky decomposition. Part of our
work was to port the RAF::Solver to Linux, containerize, and deploy it as a cloud service.
Then, our methodical profiling and benchmark analysis showed that each RAF::Solver
instance has different parallelization speedup margins and thus CPU demands, due to
each building's unique properties. Based on these observations we implemented both
static and elastic CPU-time allocation schemes. Our evaluation analysis indicates that our
fine-grained, vertically elastic CPU-time allocator yields better parallelization exploitation,
up to 77% higher resource utilization and up to x10 less SLO violations, compared to the
static allocation approaches.
|
| Language |
English |
| Subject |
Cloud computing |
|
Elasticity |
|
Ελαστικότητα |
|
Νέφος |
| Issue date |
2021-07-30 |
| Collection
|
School/Department--School of Sciences and Engineering--Department of Computer Science--Post-graduate theses
|
|
|
Type of Work--Post-graduate theses
|
| Permanent Link |
https://elocus.lib.uoc.gr//dlib/2/c/a/metadata-dlib-1625733408-968749-25145.tkl
|
| Views |
626 |
Collections
