Lock-based concurrency control suffers from programmability, scalability, and composability challenges. These challenges are exacerbated in emerging multicore architectures, on which improved software performance must be achieved by exposing greater concurrency. Transactional memory (TM) is an emerging alternative synchronization model for shared memory objects that promises to alleviate these difficulties.
In this dissertation, we consider software transactional memory (STM) for concurrency control in multicore real-time software, and present a suite of real-time STM contention managers for resolving transactional conflicts. The contention managers are called ECM, RCM, LCM, PNF, and FBLT. RCM and ECM resolve conflicts using fixed and dynamic priorities of real-time tasks, respectively, and are naturally intended to be used with the fixed priority (e.g., G-RMA) and dynamic priority (e.g., G-EDF) multicore real-time schedulers, respectively. LCM resolves conflicts based on task priorities as well as atomic section lengths, and can be used with G-EDF or G-RMA schedulers. Transactions under ECM, RCM, and LCM may retry due to conflicts with higher priority tasks even when there are no shared objects, i.e., transitive retry. PNF avoids transitive retry and optimizes processor usage by lowering the priority of retrying transactions, thereby enabling other non-conflicting transactions to proceed. PNF, however, requires a priori knowledge of all requested objects for each atomic section, which is inconsistent with the semantics of dynamic STM. Moreover, its centralized design increases overhead. FBLT avoids transitive retry, do not require a priori knowledge of requested objects, and has a decentralized design.
We establish upper bounds on transactional retry costs and task response times under the contention managers through schedulability analysis. Since ECM and RCM preserve the semantics of the underlying real-time scheduler, their maximum transactional retry cost is double the maximum atomic section length. This is improved in the design of LCM, which achieves shorter retry costs and tighter upper bounds. As PNF avoids transitive retry and improves processor usage, it yields shorter retry costs and tighter upper bounds than ECM, RCM, and LCM. FBLT\'s upper bounds are similarly tight because it combines the advantages of PNF and LCM.
We formally compare the proposed contention managers with each other, with lock-free synchronization, and with multiprocessor real-time locking protocols. Our analysis reveals that, for most cases, ECM, RCM, and LCM achieve higher schedulability than lock-free synchronization only when the atomic section length does not exceed half of lock-free synchronization\'s retry loop length. With equal periods and greater access times for shared objects, atomic section length under ECM, RCM, and LCM can be much larger than the retry loop length while still achieving better schedulability. With proper values for LCM\'s design parameters, atomic section length can be larger than the retry loop length for better schedulability. Under PNF, atomic section length can exceed lock-free\'s retry loop length and still achieve better schedulability in certain cases. FBLT achieves equal or better schedulability than lock-free with appropriate values for design parameters. The schedulability advantage of the contention managers over multiprocessor real-time locking protocols such as Global OMLP and RNLP depends upon the value of $s_{max}/L_{max}$, the ratio of the maximum transaction length to the maximum critical section length. FBLT\'s schedulability is equal or better than Global OMLP and RNLP if $s_/L_ le 2$.
Checkpointing enables partial roll-back of transactions by recording transaction execution states (i.e., checkpoints) during execution, allowing roll-back to a previous checkpoint instead of transaction start, improving task response time. We extend FBLT with checkpointing and develop CP-FBLT, and identify the conditions under which CP-FBLT achieves equal or better schedulability than FBLT.
We implement the contention managers in the Rochester STM framework and conduct experimental studies using a multicore real-time Linux kernel. Our studies reveal that among the contention managers, CP-FBLT has the best average-case performance. CP-FBLT\'s higher performance is due to the fact that PNF\'s and LCM\'s advantages are combined into the design of FBLT, which is the base of CP-FBLT. Moreover, checkpointing improves task response time. The contention managers were also found to have equal or better average-case performance than lock-free synchronization: more jobs meet their deadlines using CP-FBLT, FBLT, and PNF than lock-free synchronization by 34.6%, 28.5%, and 32.4% (on average), respectively. The superiority of the contention managers is directly due to their better conflict resolution policies.
Locking protocols such as OMLP and RNLP were found to perform better: more jobs meet their deadlines under OMLP and RNLP than any contention manager by 12.4% and 13.7% (on average), respectively. However, the proposed contention managers have numerous qualitative advantages over locking protocols. Locks do not compose, whereas STM transactions do. To allow multiple objects to be accessed in a critical section, OMLP assigns objects to non-conflicting groups, where each group is protected by a distinct lock. RNLP assumes that objects are accessed in a specific order to prevent deadlocks. In contrast, STM allows multiple objects to be accessed in a transaction in any order, while guaranteeing deadlock-freedom, which significantly increases programmability. Moreover, STM offers platform independence: the proposed contention managers can be entirely implemented in the user-space as a library. In contrast, real-time locking protocols such as OMLP and RNLP must be supported by the underlying platform (i.e., operating system or virtual machine). / Ph. D.
Identifer | oai:union.ndltd.org:VTETD/oai:vtechworks.lib.vt.edu:10919/23867 |
Date | 02 October 2013 |
Creators | El-Shambakey, Mohammed Talat |
Contributors | Electrical and Computer Engineering, Ravindran, Binoy, Rizk, Mohamed Rizk Mohamed, Patterson, Cameron D., Vullikanti, Anil Kumar S., Broadwater, Robert P. |
Publisher | Virginia Tech |
Source Sets | Virginia Tech Theses and Dissertation |
Detected Language | English |
Type | Dissertation |
Format | ETD, application/pdf |
Rights | In Copyright, http://rightsstatements.org/vocab/InC/1.0/ |
Page generated in 0.0027 seconds