DocumentCode :
3716764
Title :
Use HDLC Principles for Distributed Applications
Author :
Justin Y. Shi
Author_Institution :
Dept. of Comput. &
fYear :
2015
Firstpage :
2002
Lastpage :
2009
Abstract :
A dependable computing system should not have uncontrolled data losses. This simple requirement is easy to satisfy at the systems level but seems impossible at the application level. Due to resource sharing, this problem is more pronounced for cloud applications. It is well-known that HDLC (High-level Data Link Control) protocols can deliver reliable once-only data services using unreliable lower level services. It is not well-known that reliable communication between two crashing hosts is not implementable (the impossibility theory). Putting these two theories together, as a demonstration, this paper reports that it is possible to deliver reliable once-only services for general purpose dependable distributed applications using unreliable processors and networks under an application level tuple switched network (TSN) following the proven principles of HDLC protocols. Unlike traditional distributed programming paradigms that assume a reliable communication layer with unprotected hosts, TSN is incomplete without the application-level retransmission discipline. This application level network was designed for Zero Assumption Application Architectures (ZA3) in order to eliminate arbitrary data losses and single-point failures within TSN. The absence of single-point failure at any scale eliminates the notorious distributed applications´ scaling dilemma, making ZA3 applications suitable for extreme scale mission critical applications. Although ZA3 applications have natural fault resilience advantages, it is not clear how the ZA3 application can deliver competitive performance against "bare metal" explicit parallel programs. This paper reports computational results of a granularity tuned ZA3 wrapped matrix application compared against an unwrapped C/MPI program with fixed N/P granularity. A curious experiment using exhaustive processor/core configurations is also included in order to find the "sweet spot" for the best unwrapped C/MPI performance. Description on the implementation of ZA3 prototypes is also included.
Keywords :
"Protocols","Computer crashes","Scalability","Runtime","Reliability theory","Switches"
Publisher :
ieee
Conference_Titel :
Computer and Information Technology; Ubiquitous Computing and Communications; Dependable, Autonomic and Secure Computing; Pervasive Intelligence and Computing (CIT/IUCC/DASC/PICOM), 2015 IEEE International Conference on
Type :
conf
DOI :
10.1109/CIT/IUCC/DASC/PICOM.2015.297
Filename :
7363343
Link To Document :
بازگشت