Loading...
Please wait, while we are loading the content...
Similar Documents
Data-conscious Scheduling of Workflows in Metacomputers
| Content Provider | Semantic Scholar |
|---|---|
| Author | Lamb, P. D. Bard, Nolan Closson, Michael Ding, Meng Kan, Morgan Lee, Mark Szafron, Duane Fyshe, Alona |
| Abstract | We develop the Trellis Driver package for integrating Java applications with Trellis metacomputers, which are user-level aggregations of hosts. Using TrellisDriver.exec() calls in place of Runtime.exec() calls, applications can distribute their workflows across metacomputers. For example, Proteome Analyst (PA) is a high-performance bioinformatics tool that executes a workflow of jobs to annotate proteomes. Running all workflow jobs on a single server severely restricts throughput for large analyses. Empirical results show that Trellis Driver’s job scheduling overheads can be amortized by batching together many jobs, leading to linear speed-up of application phases. We further investigate techniques to optimize PA’s performance by reducing data movement between workflow jobs. We test our new Data-Conscious (DC) scheduling policy for Trellis in a simulation study. Simulation results show that DC scheduling is most beneficial when co-locating jobs and data offers considerable savings in either network overheads, or overheads due to application file sizes. |
| File Format | PDF HTM / HTML |
| Alternate Webpage(s) | http://webdocs.cs.ualberta.ca/~paullu/Students/nicholas.lamb.msc.thesis.dcs.pdf |
| Alternate Webpage(s) | http://www.cs.ualberta.ca/~nlamb/Writeup.pdf |
| Alternate Webpage(s) | http://www.cs.ualberta.ca/~paullu/Students/nicholas.lamb.msc.thesis.dcs.pdf |
| Language | English |
| Access Restriction | Open |
| Content Type | Text |
| Resource Type | Article |
