Online Estimation and Improvement of Cache Soft Error Vulnerability
Subject Areas : Specialمحمد معینی جهرمی 1 , محمد حسن احمدی لیوانی 2 , مصطفی ارسالی صالحی نسب 3 *
1 -
2 -
3 -
Keywords:
Abstract :
Due to the high density of transistors, memories are highly susceptible to soft errors. The processor's cache, by holding execution data and having frequent interactions with it, greatly impacts system reliability. This importance is even higher in embedded systems and safety-critical applications. One of the most significant factors affecting the reliability of the cache is its size. Smaller caches have better reliability due to their smaller area and shorter data retention, but reducing the cache size makes program execution times longer. This increases the probability of a soft error. Furthermore, reliability of cache is not uniform during program execution, and fixed size of memory cannot optimize its reliability during this time. In this regard, the main issue in improving cache vulnerability is to determine an optimum size of cache and its change time according to change overhead. Accordingly, this paper defines a model for estimating cache vulnerability, which determines vulnerability based on cache data and the type of access to it. Based on the proposed model, an algorithm has been implemented that estimates cache vulnerability online during execution. To model time in this approach, counters are used that model access times during decision-making intervals. By estimating based on blocks instead of memory words and determining the sizes of the counters and decision intervals, the proposed method has been optimized. The accuracy of the vulnerability trend estimation compared to the reference model is 95.22%. Additionally, by using the estimated vulnerability trend during execution and the effective cache size of each program, an algorithm for reconfiguring the cache to improve its vulnerability has been proposed. Implementation showed that with only 5.4% area overhead and 6% time overhead, we can have a reconfigurable memory equipped with a vulnerability management algorithm, which has a lower runtime vulnerability than a fixed cache size and overall vulnerability improvement of 36%.