DeNovo: rethinking the memory hierarchy for disciplined parallelism

shared memory;cache coherence;memory consistency;disciplined parallelism

Sung, Hyojin

关键词: shared memory;    cache coherence;    memory consistency;    disciplined parallelism;   

Others  :  https://www.ideals.illinois.edu/bitstream/handle/2142/88059/SUNG-DISSERTATION-2015.pdf?sequence=1&isAllowed=y

美国|英语

来源: The Illinois Digital Environment for Access to Learning and Scholarship

【 摘 要 】

As multicore systems become widespread, both software and hardware face a major challenge in efficiently exploiting and implementing parallelism. While shared–memory remains a popular programming model due to its global address space, it is plagued with undisciplined programming practices that allow implicit communication and unstructured non-determinism. Such “wild” shared-memory behavior not only makes it difficult to test and maintain software but also complicates hardware, preventing it from scaling in a power-efficient manner. Recent research has proposed replacing the wild shared-memory programming models with a more disciplined approach.The DeNovo project asks the following question: if software is more disciplined, can we build more power-, performance-, and complexity-efficient shared-memory hardware? Focusing on deterministic programs as a discipline to drive DeNovo, we first show that coherence and communication can be made much simpler and more efficient than the current state of the art. The resulting protocol is without transient states, invalidation traffic, directory sharer–lists, or false sharing - all significant sources of inefficiencies in existing protocols. Widening the software space further, we then show how DeNovo can support software with disciplined non-determinism without giving up its benefits for deterministic programs. The remaining challenge is supporting synchronization accesses that are inherently “racy” on DeNovo without writer-initiated invalidation. We show that arbitrary synchronization can be supported on DeNovo with a simple yet efficient hardware mechanism, a big step toward our eventual goal of supporting legacy programs. Finally, we explore the potential for a comprehensive coherence solution that merges all previous DeNovo coherence mechanisms and adaptively switches between them depending on the level of “discipline” of software.In summary, DeNovo shows the potential for commercially viable software-driven shared-memory systems with higher complexity-, performance-, and energy-efficiency than today’s software-oblivious hardware.