Temporally Bounding TSO for Fence-Free Asymmetric Synchronization

Abstract

This paper introduces a temporally bounded total store ordering (TBTSO) memory model, and shows that it enables nonblocking fence-free solutions to asymmetric synchronization problems, such as those arising in memory reclamation and biased locking.TBTSO strengthens the TSO memory model by bounding the time it takes a store to drain from the store buffer into memory. This bound enables devising fence-free algorithms for asymmetric problems, which require a performancecritical fast path to synchronize with an infrequently executed slow path. We demonstrate this by constructing (1) a fence-free version of the hazard pointers memory reclamation scheme, and (2) a fence-free biased lock algorithm which is compatible with unmanaged environments as it does not rely on safe points or similar mechanisms.We further argue that TBTSO can be implemented in hardware with modest modifications to existing TSO architectures. However, our design makes assumptions about proprietary implementation details of commercial hardware; it thus best serves as a starting point for a discussion on the feasibility of hardware TBTSO implementation. We also show how minimal OS support enables the adaptation of TBTSO algorithms to x86 systems.

Exposing errors related to weak memory in GPU applications

“…The reduced fence for sdkred-nf does not correspond to the provided fence in sdk-red; this solution may be consistent with a temporally bounded model [34] where extra instructions in one communicating thread can make up for the lack of fences in the other. The reduced fences for ls-bh-nf are a superset of the fences in ls-bh (as ls-bh showed errors with provided fences).…”

The Convergence of Stochastic Gradient Descent in Asynchronous Shared Memory

“…Intuitively, this condition holds in practice since gradients are often sparse, meaning that d is low, the delay factors τ max and τ avg are not set adversarially, and the learning rate α can be set by the user to be small enough to offset any increase in the other terms. In particular, τ max is limited by the staleness of updates in the write buffer at each core, which is well bounded in practice [21].…”

Fast and Robust Memory Reclamation for Concurrent Data Structures

“…Similarly to switching from the fast path to the fallback path, the switch in the opposite direction is signaled through setting the value of the shared fallback-flag and immediately declaring a quiescent state (lines [27][28][29][30].…”

“…Improvements on HP. Morrison et al [27] introduce a new, strengthened version of the Total Store Ordering (TSO) memory model [29] in which there is a known bound on the time it takes for writes in the store buffer to become visible in main memory. A variant of HP which does not need memory barriers is proposed.…”