The authors state in their introduction that the precise analytical reliability model for architecture (PARMA) “uses a rigorous fault generation model to address temporal multi-bit errors (MBEs), starting with the probability of SEUs [single event upsets] on a single bit and then expounding the probabilities in both temporal and spatial dimensions.”

They go on to claim,

Using the insights gained from the comparison of the PARMA model with prior approximate analytical models, we have introduced a new approximate analytical model based on a refined AVF [architectural vulnerability factor] methodology to estimate the DUE FIT rate of SECDED [single error correcting double error detecting] protected caches.

The introduction clearly explains why the topic is important, and the authors cite relevant related papers and show their contributions clearly (p. 86). Topics include goals of performance analysis and measurement, performance metrics, means, modes, measurement tools and techniques, perturbations due to measurement, and the design of experiments and simulation. The paper introduces an original and unified framework for measuring the reliability of static random-access memory (SRAM) arrays protected by any possible error protection scheme, and a new and highly accurate approximate analytical model for measuring the FIT rate of caches protected by word-level SECDED codes. The conclusions follow directly from the body of the paper. They are well structured and introduce no new material.

The technical approach is very courageous. However, for better performance on real-world applications, the authors should perhaps try in future works to cut out some limitations and confusions. First of all, the probability distribution of SEUs is assumed to be the same after every cycle. Another weakness likely to bewilder the student is the assumption that there is no correlation between SEUs affecting any two cache bits. The results show that PARMA simulation is slower than the basic sim-outorder simulation by a factor of about 25 times for 100 million SimPoint simulations. Those interested in this area of research should also study three other papers [1,2,3].