System Evaluation Of A Concurrency Control Mechanism For Distributed Shared Memory Multiprocessors

Author : Pavan Poudel
Publisher :
ISBN 13 :
Total Pages : 0 pages
Book Rating : 4.:/5 (134 download)

DOWNLOAD NOW!

Download or read book Tools and Techniques for Efficient Transactions written by Pavan Poudel and published by . This book was released on 2021 with total page 0 pages. Available in PDF, EPUB and Kindle. Book excerpt: A major challenge in modern multiprocessor computer programming is concurrency control: (i) how to coordinate accesses to memory locations shared among concurrently executing tasks and (ii) how to ensure that the computation is correct. The traditional approach is to use mutexes and locks but they have many drawbacks such as deadlocks. This dissertation explores the recently emerged paradigm of transactional memory for multiprocessor programming. In transactional memory, program code is split into transactions, blocks of code containing a sequence of read and write operations that appear to execute atomically to a set of shared resources. When transactions access the same shared resources at the same time, conflict occurs, and transactions may need to abort. A transaction commits if either no conflict occurs, or conflicts are resolved. Conflicts are typically resolved through a transaction scheduling algorithm. This dissertation explores transactional memory in the context of shared memory multiprocessor systems where concurrent tasks interact through reading and writing the same main memory as well as distributed multiprocessor systems where concurrent tasks interact by sending messages to each other. The main difference between shared memory and distributed systems is that there is non-uniformity in memory access latency in distributed systems, which is not the case in shared memory systems. This non-uniformity is vital and affects not only the total execution time of all concurrent tasks but also other related network parameters such as communication cost and congestion. For uniform latency, the focus is mainly on minimizing total execution time of all concurrent tasks. This dissertation develops several novel results in both shared memory and distributed multiprocessor systems. In shared memory systems, it develops a versioning method and a transaction scheduling mechanism. Specifically, it proposes an adaptive versioning method that switches between eager and lazy versioning methods during runtime, without knowing application characteristics a priori, so that the total execution time is always better than that is obtained applying eager and lazy versioning methods individually. This method is studied for both volatile and non-volatile shared memory systems. Additionally, an extensive experimental evaluation is performed using a state-of-the-art transactional memory system and benchmark suites and results are reported. In distributed systems, it develops a transaction scheduling method that provides an execution order of transactions minimizing total execution time and additionally related network parameter communication cost. It also develops an efficient and scalable execution framework, called GraphTM, for supporting TM in clusters and distributed systems. The proposed scheduling method is studied for (i) the data-flow execution model where transactions execute on the network node that issue them, but objects (memory) read/written by transactions move to the nodes those transactions are executing and (ii) the control-flow execution model where objects do not move but the transactions accessing the objects send requests to the node where objects are located to read/write that object. The time and communication bounds obtained are either optimal or near optimal for any offline, online, or dynamic arrival of transactions. Additionally, an extensive experimental evaluation is performed using state-of-the-art benchmark suites and results are reported.