• Refine Query
  • Source
  • Publication year
  • to
  • Language
  • No language data
  • Tagged with
  • 1
  • 1
  • 1
  • 1
  • 1
  • 1
  • 1
  • 1
  • 1
  • 1
  • 1
  • 1
  • 1
  • 1
  • 1
  • About
  • The Global ETD Search service is a free service for researchers to find electronic theses and dissertations. This service is provided by the Networked Digital Library of Theses and Dissertations.
    Our metadata is collected from universities around the world. If you manage a university/consortium/country archive and want to be added, details can be found on the NDLTD website.
1

On Optimizing Transactional Memory: Transaction Splitting, Scheduling, Fine-grained Fallback, and NUMA Optimization

Mohamedin, Mohamed Ahmed Mahmoud 01 September 2015 (has links)
The industrial shift from single core processors to multi-core ones introduced many challenges. Among them, a program cannot get a free performance boost by just upgrading to a new hardware because new chips include more processing units but at the same (or comparable) clock speed as the previous generation. In order to effectively exploit the new available hardware and thus gain performance, a program should maximize parallelism. Unfortunately, parallel programming poses several challenges, especially when synchronization is involved because parallel threads need to access the same shared data. Locks are the standard synchronization mechanism but gaining performance using locks is difficult for a non-expert programmers and without deeply knowing the application logic. A new, easier, synchronization abstraction is therefore required and Transactional Memory (TM) is the concrete candidate. TM is a new programming paradigm that simplifies the implementation of synchronization. The programmer just defines atomic parts of the code and the underlying TM system handles the required synchronization, optimistically. In the past decade, TM researchers worked extensively to improve TM-based systems. Most of the work has been dedicated to Software TM (or STM) as it does not requires special transactional hardware supports. Very recently (in the past two years), those hardware supports have become commercially available as commodity processors, thus a large number of customers can finally take advantage of them. Hardware TM (or HTM) provides the potential to obtain the best performance of any TM-based systems, but current HTM systems are best-effort, thus transactions are not guaranteed to commit in any case. In fact, HTM transactions are limited in size and time as well as prone to livelock at high contention levels. Another challenge posed by the current multi-core hardware platforms is their internal architecture used for interfacing with the main memory. Specifically, when the common computer deployment changed from having a single processor to having multiple multi-core processors, the architects redesigned also the hardware subsystem that manages the memory access from the one providing a Uniform Memory Access (UMA), where the latency needed to fetch a memory location is the same independently from the specific core where the thread executes on, to the current one with a Non-Uniform Memory Access (NUMA), where such a latency differs according to the core used and the memory socket accessed. This switch in technology has an implication on the performance of concurrent applications. In fact, the building blocks commonly used for designing concurrent algorithms under the assumptions of UMA (e.g., relying on centralized meta-data) may not provide the same high performance and scalability when deployed on NUMA-based architectures. In this dissertation, we tackle the performance and scalability challenges of multi-core architectures by providing three solutions for increasing performance using HTM (i.e., Part-HTM, Octonauts, and Precise-TM), and one solution for solving the scalability issues provided by NUMA-architectures (i.e., Nemo). • Part-HTM is the first hybrid transactional memory protocol that solves the problem of transactions aborted due to the resource limitations (space/time) of current best-effort HTM. The basic idea of Part-HTM is to partition those transactions into multiple sub-transactions, which can likely be committed in hardware. Due to the eager nature of HTM, we designed a low-overhead software framework to preserve transaction's correctness (with and without opacity) and isolation. Part-HTM is efficient: our evaluation study confirms that its performance is the best in all tested cases, except for those where HTM cannot be outperformed. However, in such a workload, Part-HTM still performs better than all other software and hybrid competitors. • Octonauts tackles the live-lock problem of HTM at high contention level. HTM lacks of advanced contention management (CM) policies. Octonauts is an HTM-aware scheduler that orchestrates conflicting transactions. It uses a priori knowledge of transactions' working-set to prevent the activation of conflicting transactions, simultaneously. Octonauts also accommodates both HTM and STM with minimal overhead by exploiting adaptivity. Based on the transaction's size, time, and irrevocable calls (e.g., system call) Octonauts selects the best path among HTM, STM, or global locking. Results show a performance improvement up to 60% when Octonauts is deployed in comparison with pure HTM with falling back to global locking. • Precise-TM is a unique approach to solve the granularity of the software fallback path of best-efforts HTM. It provide an efficient and precise technique for HTM-STM communication such that HTM is not interfered by concurrent STM transactions. In addition, the added overhead is marginal in terms of space or execution time. Precise-TM uses address-embedded locks (pointers bit-stealing) for a precise communication between STM and HTM. Results show that our precise fine-grained locking pays off as it allows more concurrency between hardware and software transactions. Specifically, it gains up to 5x over the default HTM implementation with a single global lock as fallback path. • Nemo is a new STM algorithm that ensures high and scalable performance when an application workload with a data locality property is deployed. Existing STM algorithms rely on centralized shared meta-data (e.g., a global timestamp) to synchronize concurrent accesses, but in such a workload, this scheme may hamper the achievement of scalable performance given the high latency introduced by NUMA architectures for updating those centralized meta-data. Nemo overcomes these limitations by allowing only those transactions that actually conflict with each other to perform inter-socket communication. As a result, if two transactions are non-conflicting, they cannot interact with each other through any meta-data. Such a policy does not apply for application threads running in the same socket. In fact, they are allowed to share any meta-data even if they execute non-conflicting operations because, supported by our evaluation study, we found that the local processing happening inside one socket does not interfere with the work done by parallel threads executing on other sockets. Nemo's evaluation study shows improvement over state-of-the-art TM algorithms by as much as 65%. / Ph. D.

Page generated in 0.1188 seconds