Safety analysis of the Therac-25

The Therac-25 safety analysis included (1) failure mode and effect analysis, (2) fault-tree analysis, and (3) software examination.

Failure mode and effect analysis. An FMEA describes the associated system response to all failure modes of the individual system components, considered one by one. When software was involved, AECL made no assessment of the "how and why" of software faults and took any combination of software faults as a single event. The latter means that if the software was the initiating event, then no credit was given for the software mitigating the effects. This seems like a reasonable and conservative approach to handling software faults.

Fault-tree analysis. An FMEA identifies single failures leading to Class I hazards. To identify multiple failures and quantify the results, AECL used fault-tree analysis. An FTA starts with a postulated hazard — for example, two of the top events for the Therac-25 are high dose per pulse and illegal gantry motion. The immediate causes for the event are then generated in an AND/OR tree format, using a basic understanding of the machine operation to determine the causes. The tree generation continues until all branches end in "basic events." Operationally, a basic event is sometimes defined as an event that can be quantified (for example, a resistor fails open).

AECL used a "generic failure rate" of 10-4 per hour for software events. The company justified this number as based on the historical performance of the Therac-25 software. The final report on the safety analysis said that many fault trees for the Therac-25 have a computer malfunction as a causative event, and the outcome of quantification is therefore dependent on the failure rate chosen for software.

Leaving aside the general question of whether such failure rates are meaningful or measurable for software in general, it seems rather difficult to justify a single figure of this sort for every type of software error or software behavior. It would be equivalent to assigning the same failure rate to every type of failure of a car, no matter what particular failure is considered.

The authors of the safety study did note that despite the uncertainty that software introduces into quantification, fault-tree analysis provides valuable information in showing single and multiple failure paths and the relative importance of different failure mechanisms. This is certainly true.

Software examination. Because of the difficulty of quantifying software behavior, AECL contracted for a detailed code inspection to "obtain more information on which to base decisions." The software functions selected for examination were those related to the Class I software hazards identified in the FMEA: electron-beam scanning, energy selection, beam shutoff, and dose calibration.

The outside consultant who performed the inspection included a detailed examination of each function’s implementation, a search for coding errors, and a qualitative assessment of its reliability. The consultant recommended program changes to correct shortcomings, improve reliability, or improve the software package in a general sense. The final safety report gives no information about whether any particular methodology or tools were used in the software inspection or whether someone just read the code looking for errors.

Conclusions of the safety analysis. The final report summarizes the conclusions of the safety analysis:

The conclusions of the analysis call for 10 changes to Therac-25 hardware; the most significant of these are interlocks to back up software control of both electron scanning and beam energy selection.

Although it is not considered necessary or advisable to rewrite the entire Therac-25 software package, considerable effort is being expended to update it. The changes recommended have several distinct objectives: improve the protec-tion it provides against hardware failures; provide additional reliability via cross-checking; and provide a more maintainable source package. Two or three software releases are anticipated before these changes are completed.

The implementation of these improvements including design and testing for both hardware and software is well under way. All hardware modifications should be completed and installed by mid 1989, with final software updates extending into late 1989 or early 1990.

The recommended hardware changes appear to add protection against software errors, to add extra protection against hardware failures, or to increase safety margins. The software conclusions included the following:

The software code for Beam Shut-Off, Symmetry Control, and Dose Calibration was found to be straight-forward and no execution path could be found which would cause them to perform incorrectly. A few improvements are being incor-porated, but no additional hardware interlocks are required.

Inspection of the Scanning and Energy Selection func-tions, which are under software control, showed no improper execution paths; however, software inspection was unable to provide a high level of confidence in their reliability. This was due to the complex nature of the code, the extensive use of variables, and the time limitations of the inspection process. Due to these factors and the possible clinical consequences of a malfunction, computer-independent interlocks are being retrofitted for these two cases.

Given the complex nature of this software design and the basic multitasking design, it is difficult to understand how any part of the code could be labeled "straightforward" or how confidence could be achieved that "no execution paths" exist for particular types of software behavior. However, it does appear that a conservative approach — including computer-independent interlocks — was taken in most cases. Furthermore, few examples of such safety analyses of software exist in the literature. One such software analysis was performed in 1989 on the shutdown software of a nuclear power plant, which was written by a different division of AECL.[1] Much still needs to be learned about how to perform a software-safety analysis.

Reference

1. W.C. Bowman et al., "An Application of Fault Tree Analysis to Safety-Critical Software at Ontario Hydro," Conf. Probabilistic Safety Assessment and Management, 1991.