Exploiting commutativity to reduce the cost of updates to shared data in cache-coherent systems

Abstract

We present Coup, a technique to lower the cost of updates to shared data in cache-coherent systems. Coup exploits the insight that many update operations, such as additions and bitwise logical operations, are commutative: they produce the same final result regardless of the order they are performed in. Coup allows multiple private caches to simultaneously hold update-only permission to the same cache line. Caches with updateonly permission can locally buffer and coalesce updates to the line, but cannot satisfy read requests. Upon a read request, Coup reduces the partial updates buffered in private caches to produce the final value. Coup integrates seamlessly into existing coherence protocols, requires inexpensive hardware, and does not affect the memory consistency model.We apply Coup to speed up single-word updates to shared data. On a simulated 128-core, 8-socket system, Coup accelerates state-of-the-art implementations of update-heavy algorithms by up to 2.4×.

Exploiting semantic commutativity in hardware speculation

“…While we have not verified COMMTM's protocol extensions, they are similar to Coup's, which has reasonable verification complexity (by merging S and U, Coup requires no extra stable states and only 1-5 transient states [54]). …”

“…The reason is that they rely on the coherence protocol, which operates in terms of reads and writes. Recently, Coup [54] has shown that the coherence protocol can be extended to support local and concurrent commutative updates. Coup allows multiple caches to simultaneously hold update-only permission to the same cache line.…”

“…COMMTM's U state is similar to Coup's update-only state [54]. However, COMMTM requires substantially different support from Coup in nearly all other aspects: whereas Coup requires new update-only instructions for each commutative operation, COMMTM allows programs to implement arbitrary commutative operations, exploiting transactional memory to make them atomic; whereas Coup implements fixed-function reduction units, COMMTM allows arbitrary reduction functions; and whereas Coup focuses on reducing communication in a nontransactional context, COMMTM reduces both transactional conflicts and communication.…”

“…Other protocols: While we have used MSI for simplicity, COMMTM can easily extend other invalidation-based protocols, such as MESI or MOESI, with the U state [54]. In fact, we use and extend MESI in our evaluation.…”

Aim

“…The concept of merging aggregation in caches has also been used to improve the scalability of parallel reduction. Although such a benefit is completely orthogonal to our work, we compare AIM against two parallel in-cache reduction mechanisms: PCLR [Garzarán et al 2001] and Coup [Zhang et al 2015]. Both techniques alleviate communication overheads of reduction by letting each core locally perform reduction on its private cache blocks, but unlike our work, they do not explore the energy perspective of reduction.…”