Experimental evaluation of the impact of processor faults on parallel applications

This paper addresses the problem of processor faults in distributed memory parallel systems. It shows that transient faults injected at the processor pins of one node of a commercial parallel computer, without any particular fault-tolerant techniques, can cause erroneous application results for up to 43% of the injected faults (depending on the application). In addition to these very subtle faults, up to 19% of the injected faults (almost independent on the application) caused the system to hang up. These results show that fault-tolerant techniques are absolutely required in parallel systems, not only to ensure the completion of long-run applications but, and more important, to achieve confidence in the application results. The benefits of including some fairly simple behaviour based error detection mechanisms in the system were evaluated together with Algorithm Based Fault Tolerance (ABFT) techniques. The inclusion of such Mechanisms in parallel systems seems to be very important for detecting most of those subtle errors without greatly affecting the performance and the cost of these systems