Assessing the effects of communication faults on parallel applications

Abstract

This paper addresses the problem of injection of faults in the communication system of disjoint menuuy parallel computers and presents .fuult injection resultsshowing that 5% to 3Wo of the fuults injected in the conzmunication subsystem of a commercial purullel computer caused undetected errors thut l e d the upplicution to generate erroneous results. All these C U .~ correspond to situations in which it would be virtuidly inzpossible to detect that the benchmark output wus erroneous, us the size of the results file wus pluusible and no system errors h d been detected. This emphasises the n e d for .fuult tolerant techniques in parallel systems in order to uchieve confidence in the application results. This is especially true in massively parallel computers, us the probability of occurring faults increase with the number of processing nodes. Moreover, in disjoint memory computers, which is the most popular and scaluble parallel architecture, the conzmunication subsystem plays an importunt role, and is also very prone to errors. CSFI (Conznzunicution Sofiwure Fuult Injector) is a versutile tool to inject conznzunicution ,fiiults in purallel computers. Fuults injecttd with CSFI directly emulate communication faults und spurious messuges generated by non fail-silent nodes by sojbvur-e, allowing the evaluation of the impact of.fuults in purullel system, urd the assessment of fault tolerant techniques. The use o f CSFI is nearly trunspurent to the target upplicution us it only requires minor uduptutions. Deterministic ,fuults o f different nuture can be injected without user intervention and fuult injection results ure collectd uutonzuticully by CSFI.

Injecting Inconsistent Values Caused by Interaction Faults for Experimental Dependability Evaluation

“…We do not consider faults caused by transmission errors in the physical layer -injection for this type of communication fault is already addressed by previous work such as [8,3,9,14]. That means messages are not corrupted or lost in the communication channel, but are indeed generated containing incorrect data or are not sent.…”

“…Fault injection has been used in several domains, such as fault-tolerant systems [15,13], parallel computers [3,14], distributed systems [8,9] and databases [19]. Different types of faults have been addressed, including hardware [15], software [12,16,17,13], operator [4,19] and even communication faults in a more general sense [8,3,9,14].…”

“…Different types of faults have been addressed, including hardware [15], software [12,16,17,13], operator [4,19] and even communication faults in a more general sense [8,3,9,14]. Even though most existing tools focus on specific domains (e.g.…”

“…Some work also address communication faults in parallel or distributed systems [3,8,9,14], usually aiming at transmission errors in the transport or lower layers. In this paper, we also consider networked systems, but address the injection of interaction faults -i.e.…”

“…For that reason, we focus on a specific type of error: inconsistent values, data that is correctly received and syntactically correct, but inconsistent with what it should represent. Though existing approaches are able to inject data errors [8,3,17,14], we focus on a different scenario and, more important, on a different level of abstraction, where both the syntax and the semantics (meaning) of the affected data are taken into account.…”