Analyzing heap error behavior in embedded JVM environments

Recent studies have shown that transient hardware errors caused by external factors such as alpha particles and cosmic ray strikes can be responsible for a large percentage of system down-time. Denser processing technologies, increasing clock speeds, and low supply voltages used in embedded systems can worsen this problem. In many embedded environments, one may not want to provision extensive error protection in hardware because of (i) form-factor or power consumption limitations, and/or (ii) to keep costs low. Also, the mismatch between the hardware protection granularity and the field access granularity can lead to false alarms and error cancellations. Consequently, software-based approaches to identify and possibly rectify these errors seem to be promising. Towards this goal, This work specifically looks to enhance the software's ability to detect heap memory errors in a Java-based embedded system. Using several embedded Java applications, This work first studies the tradeoffs between reliability, performance, and memory space overhead for two schemes that perform error checks at object and field granularities. We also study the impact of object characteristics (e.g., lifetime, re-use intervals, access frequency, etc.) on error propagation. Considering the pros and cons of these two schemes, we then investigate two hybrid strategies that attempt to strike a balance between memory space and performance overheads and reliability. Our experimental results clearly show that the granularity of error protection and its frequency can significantly impact static/dynamic overheads and error detection ability.