Dmitri Perelman

On maintaining multiple versions in STM

An effective way to reduce the number of aborts in software transactional memory (STM) is to keep multiple versions of transactional objects. In this paper, we study inherent properties of STMs that use multiple versions to guarantee successful commits of all read-only transactions. We first show that these STMs cannot be disjoint-access parallel. We then consider the problem of garbage collecting old object versions, and show that no STM can be optimal in the number of previous versions kept. Moreover, we show that garbage collecting useless versions is impossible in STMs that implement invisible reads. Finally, we present an STM algorithm using visible reads that efficiently garbage collects useless object versions.

SMV: selective multi-versioning STM

We present Selective Multi-Versioning (SMV), a new STM that reduces the number of aborts, especially those of long read-only transactions. SMV keeps old object versions as long as they might be useful for some transaction to read. It is able to do so while still allowing reading transactions to be invisible by relying on automatic garbage collection to dispose of obsolete versions. SMV is most suitable for read-dominated workloads, for which it performs better than previous solutions. It has an up to ×7 throughput improvement over a single-version STMand more than a two-fold improvement over an STMkeeping a constant number of versions per object. We show that the memory consumption of algorithms keeping a constant number of versions per object might grow exponentially with the number of objects, while SMV operates successfully even in systems with stringent memory constraints.

SALSA: scalable and low synchronization NUMA-aware algorithm for producer-consumer pools

We present a highly-scalable non-blocking producer-consumer task pool, designed with a special emphasis on lightweight synchronization and data locality. The core building block of our pool is SALSA, Scalable And Low Synchronization Algorithm for a single-consumer container with task stealing support. Each consumer operates on its own SALSA container, stealing tasks from other containers if necessary. We implement an elegant self-tuning policy for task insertion, which does not push tasks to overloaded SALSA containers, thus decreasing the likelihood of stealing. SALSA manages large chunks of tasks, which improves locality and facilitates stealing. SALSA uses a novel approach for coordination among consumers, without strong atomic operations or memory barriers in the fast path. It invokes only two CAS operations during a chunk steal. Our evaluation demonstrates that a pool built using SALSA containers scales linearly with the number of threads and significantly outperforms other FIFO and non-FIFO alternatives.

Multi-versioning in Transactional Memory

Reducing the number of aborts is one of the biggest challenges of most transactional systems: existing TMs may abort many transactions that could, in fact, commit without violating correctness. Historically, the commonly used method for reducing the abort rate was maintaining multiple object versions. Multiversion concurrency control is a classical approach for providing concurrent access to the database in database management systems. Its idea is to let a reading transaction obtain a consistent snapshot corresponding to an arbitrary point in time (e.g., defined at the beginning of a transaction) – concurrent updates are isolated through maintaining old versions rather than via scheduling decisions.

Brief announcement: reconfigurable state machine replication from non-reconfigurable building blocks

Reconfigurable state machine replication is an important enabler of elasticity for replicated cloud services, which must be able to dynamically adjust their size as a function of changing load and resource availability. We introduce a new generic framework to allow the reconfigurable state machine implementation to be derived from a collection of arbitrary non-reconfigurable state machines. Our reduction framework follows the black box approach, and does not make any assumptions with respect to its execution environment apart from reliable channels. It allows higher-level services to leverage speculative command execution to ensure uninterrupted progress during the reconfiguration periods as well as in situations where failures prevent the reconfiguration agreement from being reached in a timely fashion. We apply our framework to obtain a reconfigurable speculative state machine from the non-reconfigurable Paxos implementation, and analyze its performance on a realistic distributed testbed. Our results show that our framework incurs negligible overheads in the absence of reconfiguration, and allows steady throughput to be maintained throughout the reconfiguration periods.

Lightweight automatic face annotation in media pages

Labeling human faces in images contained in Web media stories enables enriching the user experience offered by media sites. We propose a lightweight framework for automatic image annotation that exploits named entities mentioned in the article to significantly boost the accuracy of face recognition. While previous works in the area labor to train comprehensive offline visual models for a pre-defined universe of candidates, our approach models the people mentioned in a given story on the y, using a standard Web image search engine as an image sampling mechanism. We overcome multiple sources of noise introduced by this ad-hoc process, to build a fast and robust end-to-end system from off-the-shelf error-prone text analysis and machine vision components. In experiments conducted on approximately 900 faces depicted in 500 stories from a major celebrity news website, we were able to correctly label 81.5% of the faces while mislabeling 14.8% of them.

On Avoiding Spare Aborts in Transactional Memory

This paper takes a step toward developing a theory for understanding aborts in transactional memory systems (TMs). Existing TMs may abort many transactions that could, in fact, commit without violating correctness. We call such unnecessary aborts spare aborts. We classify what kinds of spare aborts can be eliminated, and which cannot. We further study what kinds of spare aborts can be avoided efficiently. Specifically, we show that some spare aborts cannot be avoided, and that there is an inherent tradeoff between the overhead of a TM and the extent to which it reduces the number of spare aborts. We also present an efficient example TM algorithm that avoids certain kinds of spare aborts, and analyze its properties and performance.