![Institutional Repository [SANDBOX]](images/dias.logo.png "Institutional Repository [SANDBOX]"){kind=logo}
Institutional Repository [SANDBOX]
Technical University of Crete
Technical University of Crete
EN
|
EL
| URI: | http://purl.tuc.gr/dl/dias/6A2C4F40-98E1-4C6E-BA21-E0D37A076DFC | ||
| Year | 2015 | ||
| Type of Item | Master Thesis | ||
| License |
|
||
| Bibliographic Citation | Charalabos Vatsolakis, "Α user-transparent system for virtualizing reconfigurable hardware accelerators", Master Thesis, School of Electronic and Computer Engineering, Technical University of Crete, Chania, Greece, 2015 https://doi.org/10.26233/heallink.tuc.62753 | ||
| Appears in Collections |
Hardware based acceleration is highly efficient, but there are several factors limiting its adoption. The most notable of those, is the lack of a standardized system, capable of providing a transparent interface between software and reconfigurable hardware. The product of this work, is a system capable of loading acceletors and perform I/Os in a completely transparent to the user manner. The user is capable of implementing an accelerator compatible to the system by using a standard set of ports. The access to this accelerator is aided by a given software API.The system is based on the PCI Express interface (version 1, 4 lanes) for data transactions and the ICAP for reconfiguration. There are three partially reconfigurable regions available, while the systems software is responsible for scheduling the accelerators waiting for execution. There are four scheduling policies implemented; noop, simple, out of order, and forced. The first two, take in account the submission order, while the others, target to reduce the number of reconfigurations. The total throughput our system reached, equals 618 MB/s for transmissions, 544 MB/s for receptions and 488 MB/s for reconfigurations. The behaviour of our system under heavy load, was evaluated through an edge detection system consisting of four accelerators running sequentially. We sustained a real-time throughput of 30 FPS in 720p HD datasets, even when the least efficient scheduling policy was selected.
Copyright © DIAS 2013
