University of Patras
Reduction of the average-performance degradation and/or performance variance caused by defects in CPU caches
Abstract
dc:descriptionProgress in integration technologies has played a crucial role in the evolution of integrated circuits, which consist of a large and increasingly growing number of transistors. On the other hand, as we progress towards smaller integration technologies, the number of defective elements significantly increases. These defective elements include transistors or memory structures that do not operate according to the initial specifications. Modern studies have shown that the probability of having defective transistors increases significantly as integration technology approaches the limits of the laws of physics for the construction of electronic components using silicon. This dissertation proposes methodologies for the reduction of the performance degradation and performance variance caused by faults in the memory structures of CPU caches. The focus is on memory elements for two main reasons. Firstly, memory structures occupy a very large portion of the processor (this percentage increases as we move towards multi-core processing systems). The second reason is memory structures are implemented with minimal-sized transistors, which are more sensitive to process variations. Increasing demands for reducing power consumption of integrated circuits lead to a growing need to reduce the supply voltage in combination with a decrease in operating frequency. The supply voltage has a significant impact on reducing dynamic power due to the square term in the relationship that connects them. However, a drop in voltage below the "safe operating" limits significantly contributes to the occurrence of faults. Additionally, two factors that cumulatively contribute to the number of malfunctioning circuits are their aging and external factors, such as cosmic radiation. In the first part of the dissertation, an overview of the importance of fault tolerance in the cache memory subsystem of a processing system is provided, and the basic methodologies for addressing the problem are presented. More specifically, in Chapter 1, a brief description of the characteristics of cache memory structures is given, and the reliability issue is introduced. Then, we present the methodologies that have been proposed in the literature aimed at addressing the problem of reliability of cache memories and classify them into two main categories. Finally, the problem of the presence of faults in cache memory structures is reduced to the problem of performance degradation, as well as the problem of performance variance (unpredictability). These two problems are the main reason for the imperative need for methodologies to address them and constitute the motivation for the doctoral dissertation. In Chapter 2, we specify the problem in the presence of memory elements that do not operate according to their specifications. A presentation of fault models is made, which are used to describe the characteristics and the nature of faults in integrated circuits. Following that, a reference is made to the number, as well as to the dispersion of the malfunctioning memory elements as the technology scaling continues, as well as the level of supply voltage reduces (near to the threshold voltage of transistors). Finally, this section presents the basic concepts that will be used in the continuation of the dissertation. Chapter 3 presents the simulation environment and the necessary extensions developed to evaluate the proposed methodologies. The tools and techniques used for the implementation and performance evaluation of the proposed methodologies, as well as the reliability of the results, are introduced. Two simulation environments are presented, and the advantages and disadvantages of each are evaluated. The first environment is based on a widely used architectural simulator, expanded to support the simulation of memory structures containing faulty elements. The second environment was developed entirely for the purposes of this dissertation and is based on reading memory system accesses from a file for study. Closing this chapter, the benchmark suites used for evaluating the proposed methodologies are presented. In the following sections of the dissertation, we present the proposed methodologies developed as part of the doctoral research. In Chapter 4, we introduce a fault tolerance methodology based on the use of error detection and correction codes. Specifically, it analyzes how the problem of using error detection and correction codes introduces significant delays during error correction, rendering them not suitable for time-sensitive first-level caches. This delay is proportional to the number of errors that the code needs to correct. The proposed methodology attempts to predict the criticality of the instruction to be executed. The prediction information has dual significance: in the case of a cache miss in the first-level cache memory, this prediction affects the replacement algorithm, while in the case where the data is found in the cache memory, the prediction of the instruction's criticality determines the error detection and correction code to be applied. Chapter 5, a methodology is described that relies on the use of an additional very small memory that operates collaboratively with the first-level cache memory. This new memory serves as an agent recovering the performance loss due to disabled faulty blocks of cache memory (spare cache) and also captures the evicted blocks in order to reduce conflict misses of the cache memory (victim-cache).Chapter 6 describes a new methodology based on predicting the spatial locality of the memory blocks introduced into the first-level cache memories. A comprehensive evaluation of memory reference spatial locality prediction mechanisms is performed, leading to the design of a new mechanism with minimal hardware requirements that maintains the correctness and reliability of predictions. This methodology benefits from improved management of the fault-free cache memory utilization, enhancing the input and replacement algorithm of cache memories by considering the spatial locality of memory references. In Chapter 7, a methodology is presented that utilizes additional hardware and has the ability to statically adapt the size of the first-level cache memory block based on the number of faults. The block size changes in coordination with the cache associativity. This ensures that the cache's data structure remains unchanged, and the necessary modifications to the underlying cache architecture are minimized. Closing Chapter 7, the question of the most appropriate type of redundancy (additional storage area) for storing data or control based information (extra tag and status bits) is analyzed. In Chapters 8 and 9, two additional methodologies are described, based on the different behavior of the phases of the executed program concerning the presence of faults in the memory subsystem. The problem is elevated not only to the level of the number of faults but also studies the behavior of different applications and different program phases of the same application concerning the presence of faults. The two presented methodologies share the element of periodic real-time control and adaptation of the organization of the cache memory to increase the system's performance. The two proposed evaluation and reconfiguration mechanisms, based on time window division, are presented. Additionally, an analysis is conducted on how to reduce costly (in terms of performance and consumed energy) actions that need to be performed before each reconfiguration of the first-level cache memories. In Chapter 9, the problem of predicting the execution time of a program is examined when running on a processing system in which the memory subsystem has faulty cells. In systems executing critical applications, there are strictly defined time constraints that must be met by specific processes. As demonstrated, the presence of faults in cache memories significantly contributes to the uncertainty of the time needed to complete an application. The methodology presented in this chapter aims to increase the predictability of execution time while simultaneously reducing the impact of faults on the performance of the memory subsystem. As a result, this methodology manages to improve the overall quality of service delivered to the end user. Finally, Chapter 10 concludes the dissertation by providing general conclusions and directions for future research. It also includes the publications based on the research results of the dissertation and the citations.
Degree
thesis:*- Grantor dc:publisher
- University of Patras
- Year dc:date
- 2024
Author and committee
dc:creator, dc:contributor.*- Authors dc:creator
-
- Mavropoulos, Michail
- Μαυρόπουλος, Μιχαήλ
Subjects
dc:subject × 14- Αρχιτεκτονική κρυφών μνημών
- Ανοχή σφαλμάτων
- Βαθμιαία υποβάθμιση της απόδοσης
- Αξιοπιστία
- Cache architecture
- Fault tolerance
- Graceful degradation
- Reliability
- Επιστήμη Ηλεκτρολόγου Μηχανικού, Ηλεκτρονικού Μηχανικού, Μηχανικού Η/Υ
- Επιστήμες Μηχανικού και Τεχνολογία
- Υπολογιστές, Υλικό (hardware) και Αρχιτεκτονική
- Electrical Engineering, Electronic Engineering, Information Engineering
- Engineering and Technology
- Computer science, Hardware and Architecture
Rights
- Language dc:language
- eng
Identifiers
dc:identifier.*- Identifier
- 10.12681/eadd/57629