arXiv subjects

Petr Kuznetsov

Publications and source records attributed to Petr Kuznetsov.

At least 55 records · Page 3Linked to original sources

arXiv2017

A Concurrency-Optimal Binary Search Tree

The paper presents the first \emph{concurrency-optimal} implementation of a binary search tree (BST). The implementation, based on a standard sequential implementation of an internal tree, ensures that every \emph{schedule} is accepted, i.e., interleaving of steps of the sequential code, unless linearizability is violated. To ensure this property, we use a novel read-write locking scheme that protects tree \emph{edges} in addition to nodes. Our implementation outperforms the state-of-the art BSTs on most basic workloads, which suggests that optimizing the set of accepted schedules of the sequential code can be an adequate design principle for efficient concurrent data structures.

cs.DC↗

arXiv2016

Set-Consensus Collections are Decidable

A natural way to measure the power of a distributed-computing model is to characterize the set of tasks that can be solved in it. %the model. In general, however, the question of whether a given task can be solved in a given model is undecidable, even if we only consider the wait-free shared-memory model. In this paper, we address this question for restricted classes of models and tasks. We show that the question of whether a collection $C$ of \emph{$(\ell,j)$-set consensus} objects, for various $\ell$ (the number of processes that can invoke the object) and $j$ (the number of distinct outputs the object returns), can be used by $n$ processes to solve wait-free $k$-set consensus is decidable. Moreover, we provide a simple $O(n^2)$ decision algorithm, based on a dynamic programming solution to the Knapsack optimization problem. We then present an \emph{adaptive} wait-free set-consensus algorithm that, for each set of participating processes, achieves the best level of agreement that is possible to achieve using $C$. Overall, this gives us a complete characterization of a read-write model defined by a collection of set-consensus objects through its \emph{set-consensus power}.

cs.DC↗

arXiv2016

Read-Write Memory and k-Set Consensus as an Affine Task

The wait-free read-write memory model has been characterized as an iterated \emph{Immediate Snapshot} (IS) task. The IS task is \emph{affine}---it can be defined as a (sub)set of simplices of the standard chromatic subdivision. It is known that the task of \emph{Weak Symmetry Breaking} (WSB) cannot be represented as an affine task. In this paper, we highlight the phenomenon of a "natural" model that can be captured by an iterated affine task and, thus, by a subset of runs of the iterated immediate snapshot model. We show that the read-write memory model in which, additionally, $k$-set-consensus objects can be used is, unlike WSB, "natural" by presenting the corresponding simple affine task captured by a subset of $2$-round IS runs. Our results imply the first combinatorial characterization of models equipped with abstractions other than read-write memory that applies to generic tasks.

cs.DC↗

arXiv2016

Why Transactional Memory Should Not Be Obstruction-Free

Transactional memory (TM) is an inherently optimistic abstraction: it allows concurrent processes to execute sequences of shared-data accesses (transactions) speculatively, with an option of aborting them in the future. Early TM designs avoided using locks and relied on non-blocking synchronization to ensure obstruction-freedom: a transaction that encounters no step contention is not allowed to abort. However, it was later observed that obstruction-free TMs perform poorly and, as a result, state-of-the-art TM implementations are nowadays blocking, allowing aborts because of data conflicts rather than step contention. In this paper, we explain this shift in the TM practice theoretically, via complexity bounds. We prove a few important lower bounds on obstruction-free TMs. Then we present a lock-based TM implementation that beats all of these lower bounds. In sum, our results exhibit a considerable complexity gap between non-blocking and blocking TM implementations.

cs.DC↗

arXiv2016

In the Search of Optimal Concurrency

Implementing a concurrent data structure typically begins with defining its sequential specification. However, when used \emph{as is}, a nontrivial sequential data structure, such as a linked list, a search tree, or a hash table, may expose incorrect behavior: lost updates, inconsistent responses, etc. To ensure correctness, portions of the sequential code operating on the shared data must be "protected" from data races using synchronization primitives and, thus, certain schedules of the steps of concurrent operations must be rejected. But can we ensure that we do not "overuse" synchronization, i.e., that we reject a concurrent schedule only if it violates correctness? In this paper, we treat this question formally by introducing the notion of a \emph{concurrency-optimal} implementation. A program's concurrency is defined here as its ability to accept concurrent schedules, i.e., interleavings of steps of its sequential implementation. An implementation is concurrency-optimal if it accepts all interleavings that do not violate the program's correctness. We explore the concurrency properties of \emph{search} data structures which can be represented in the form of directed acyclic graphs exporting insert, delete and search operations. We prove, for the first time, that \emph{pessimistic} e.g., based on conservative locking) and \emph{optimistic serializable} e.g., based on serializable transactional memory) implementations of search data-structures are incomparable in terms of concurrency. Specifically, there exist simple interleavings of sequential code that cannot be accepted by \emph{any} pessimistic (and \emph{resp.}, serializable optimistic) implementation, but accepted by a serializable optimistic one (and \emph{resp.}, pessimistic). Thus, neither of these two implementation classes is concurrency-optimal.

cs.DC↗

arXiv2015

Progressive Transactional Memory in Time and Space

Transactional memory (TM) allows concurrent processes to organize sequences of operations on shared \emph{data items} into atomic transactions. A transaction may commit, in which case it appears to have executed sequentially or it may \emph{abort}, in which case no data item is updated. The TM programming paradigm emerged as an alternative to conventional fine-grained locking techniques, offering ease of programming and compositionality. Though typically themselves implemented using locks, TMs hide the inherent issues of lock-based synchronization behind a nice transactional programming interface. In this paper, we explore inherent time and space complexity of lock-based TMs, with a focus of the most popular class of \emph{progressive} lock-based TMs. We derive that a progressive TM might enforce a read-only transaction to perform a quadratic (in the number of the data items it reads) number of steps and access a linear number of distinct memory locations, closing the question of inherent cost of \emph{read validation} in TMs. We then show that the total number of \emph{remote memory references} (RMRs) that take place in an execution of a progressive TM in which $n$ concurrent processes perform transactions on a single data item might reach $Ω(n \log n)$, which appears to be the first RMR complexity lower bound for transactional memory.

cs.DC↗

arXiv2015

Optimism for Boosting Concurrency

Modern concurrent programming benefits from a large variety of synchronization techniques. These include conventional pessimistic locking, as well as optimistic techniques based on conditional synchronization primitives or transactional memory. Yet, it is unclear which of these approaches better leverage the concurrency inherent to multi-cores. In this paper, we compare the level of concurrency one can obtain by converting a sequential program into a concurrent one using optimistic or pessimistic techniques. To establish fair comparison of such implementations, we introduce a new correctness criterion for concurrent programs, defined independently of the synchronization techniques they use. We treat a program's concurrency as its ability to accept a concurrent schedule, a metric inspired by the theories of both databases and transactional memory. We show that pessimistic locking can provide strictly higher concurrency than transactions for some applications whereas transactions can provide strictly higher concurrency than pessimistic locks for others. Finally, we show that combining the benefits of the two synchronization techniques can provide strictly more concurrency than any of them individually. We propose a list-based set algorithm that is optimal in the sense that it accepts all correct concurrent schedules. As we show via experimentation, the optimality in terms of concurrency is reflected by scalability gains.

cs.DC↗

arXiv2015

On the Space Complexity of Set Agreement

The $k$-set agreement problem is a generalization of the classical consensus problem in which processes are permitted to output up to $k$ different input values. In a system of $n$ processes, an $m$-obstruction-free solution to the problem requires termination only in executions where the number of processes taking steps is eventually bounded by $m$. This family of progress conditions generalizes wait-freedom ($m=n$) and obstruction-freedom ($m=1$). In this paper, we prove upper and lower bounds on the number of registers required to solve $m$-obstruction-free $k$-set agreement, considering both one-shot and repeated formulations. In particular, we show that repeated $k$ set agreement can be solved using $n+2m-k$ registers and establish a nearly matching lower bound of $n+m-k$.

cs.DC↗

arXiv2015

The Weakest Failure Detector for Eventual Consistency

In its classical form, a consistent replicated service requires all replicas to witness the same evolution of the service state. Assuming a message-passing environment with a majority of correct processes, the necessary and sufficient information about failures for implementing a general state machine replication scheme ensuring consistency is captured by the Ω failure detector. This paper shows that in such a message-passing environment, Ω is also the weakest failure detector to implement an eventually consistent replicated service, where replicas are expected to agree on the evolution of the service state only after some (a priori unknown) time. In fact, we show that Ω is the weakest to implement eventual consistency in any message-passing environment, i.e., under any assumption on when and where failures might occur. Ensuring (strong) consistency in any environment requires, in addition to Ω, the quorum failure detector Σ. Our paper thus captures, for the first time, an exact computational difference be- tween building a replicated state machine that ensures consistency and one that only ensures eventual consistency.

cs.DC↗

arXiv2015

Inherent Limitations of Hybrid Transactional Memory

Several Hybrid Transactional Memory (HyTM) schemes have recently been proposed to complement the fast, but best-effort, nature of Hardware Transactional Memory (HTM) with a slow, reliable software backup. However, the fundamental limitations of building a HyTM with nontrivial concurrency between hardware and software transactions are still not well understood. In this paper, we propose a general model for HyTM implementations, which captures the ability of hardware transactions to buffer memory accesses, and allows us to formally quantify and analyze the amount of overhead (instrumentation) of a HyTM scheme. We prove the following: (1) it is impossible to build a strictly serializable HyTM implementation that has both uninstrumented reads and writes, even for weak progress guarantees, and (2) under reasonable assumptions, in any opaque progressive HyTM, a hardware transaction must incur instrumentation costs linear in the size of its data set. We further provide two upper bound implementations whose instrumentation costs are optimal with respect to their progress guarantees. In sum, this paper captures for the first time an inherent trade-off between the degree of concurrency a HyTM provides between hardware and software transactions, and the amount of instrumentation overhead the implementation must incur.

cs.DC↗

arXiv2014

Simple CHT: A New Derivation of the Weakest Failure Detector for Consensus

The paper proposes an alternative proof that Omega, an oracle that outputs a process identifier and guarantees that eventually the same correct process identifier is output at all correct processes, provides minimal information about failures for solving consensus in read-write shared-memory systems: every oracle that gives enough failure information to solve consensus can be used to implement Omega. Unlike the original proof by Chandra, Hadzilacos and Toueg (CHT), the proof presented in this paper builds upon the very fact that 2-process wait-free consensus is impossible. Also, since the oracle that is used to implement can solve consensus, the implementation is allowed to directly access consensus objects. As a result, the proposed proof is shorter and conceptually simpler than the original one.

cs.DC↗

arXiv2014

On Partial Wait-Freedom in Transactional Memory

Transactional memory (TM) is a convenient synchronization tool that allows concurrent threads to declare sequences of instructions on shared data as speculative \emph{transactions} with "all-or-nothing" semantics. It is known that dynamic transactional memory cannot provide \emph{wait-free} progress in the sense that every transaction commits in a finite number of its own steps. In this paper, we explore the costs of providing wait-freedom to only a \emph{subset} of transactions. Since most transactional workloads are believed to be read-dominated, we require that read-only transactions commit in the wait-free manner, while updating transactions are guaranteed to commit only if they run in the absence of concurrency. We show that this kind of partial wait-freedom, combined with attractive requirements like read invisibility or disjoint-access parallelism, incurs considerable complexity costs.

cs.DC↗

arXiv2014

A generalized asynchronous computability theorem

We consider the models of distributed computation defined as subsets of the runs of the iterated immediate snapshot model. Given a task $T$ and a model $M$, we provide topological conditions for $T$ to be solvable in $M$. When applied to the wait-free model, our conditions result in the celebrated Asynchronous Computability Theorem (ACT) of Herlihy and Shavit. To demonstrate the utility of our characterization, we consider a task that has been shown earlier to admit only a very complex $t$-resilient solution. In contrast, our generalized computability theorem confirms its $t$-resilient solvability in a straightforward manner.

cs.DC↗

arXiv2014

A Distributed SDN Control Plane for Consistent Policy Updates

Software-defined networking (SDN) is a novel paradigm that out-sources the control of packet-forwarding switches to a set of software controllers. The most fundamental task of these controllers is the correct implementation of the \emph{network policy}, i.e., the intended network behavior. In essence, such a policy specifies the rules by which packets must be forwarded across the network. This paper studies a distributed SDN control plane that enables \emph{concurrent} and \emph{robust} policy implementation. We introduce a formal model describing the interaction between the data plane and a distributed control plane (consisting of a collection of fault-prone controllers). Then we formulate the problem of \emph{consistent} composition of concurrent network policy updates (short: the \emph{CPC Problem}). To anticipate scenarios in which some conflicting policy updates must be rejected, we enable the composition via a natural \emph{transactional} interface with all-or-nothing semantics. We show that the ability of an $f$-resilient distributed control plane to process concurrent policy updates depends on the tag complexity, i. e., the number of policy labels (a.k.a. \emph{tags}) available to the controllers, and describe a CPC protocol with optimal tag complexity $f+2$.

cs.DC↗

arXiv2014

Strong Equivalence Relations for Iterated Models

The Iterated Immediate Snapshot model (IIS), due to its elegant geometrical representation, has become standard for applying topological reasoning to distributed computing. Its modular structure makes it easier to analyze than the more realistic (non-iterated) read-write Atomic-Snapshot memory model (AS). It is known that AS and IIS are equivalent with respect to \emph{wait-free task} computability: a distributed task is solvable in AS if and only if it solvable in IIS. We observe, however, that this equivalence is not sufficient in order to explore solvability of tasks in \emph{sub-models} of AS (i.e. proper subsets of its runs) or computability of \emph{long-lived} objects, and a stronger equivalence relation is needed. In this paper, we consider \emph{adversarial} sub-models of AS and IIS specified by the sets of processes that can be \emph{correct} in a model run. We show that AS and IIS are equivalent in a strong way: a (possibly long-lived) object is implementable in AS under a given adversary if and only if it is implementable in IIS under the same adversary. %This holds whether the object is one-shot or long-lived. Therefore, the computability of any object in shared memory under an adversarial AS scheduler can be equivalently investigated in IIS.

cs.DC↗

arXiv2013

Non-Interference and Local Correctness in Transactional Memory

Transactional memory promises to make concurrent programming tractable and efficient by allowing the user to assemble sequences of actions in atomic transactions with all-or-nothing semantics. It is believed that, by its very virtue, transactional memory must ensure that all committed transactions constitute a serial execution respecting the real-time order. In contrast, aborted or incomplete transactions should not "take effect." But what does "not taking effect" mean exactly? It seems natural to expect that aborted or incomplete transactions do not appear in the global serial execution, and, thus, no committed transaction can be affected by them. We investigate another, less obvious, feature of "not taking effect" called non-interference: aborted or incomplete transactions should not force any other transaction to abort. In the strongest form of non-interference that we explore in this paper, by removing a subset of aborted or incomplete transactions from the history, we should not be able to turn an aborted transaction into a committed one without violating the correctness criterion. We show that non-interference is, in a strict sense, not implementable with respect to the popular criterion of opacity that requires all transactions (be they committed, aborted or incomplete) to witness the same global serial execution. In contrast, when we only require local correctness, non-interference is implementable. Informally, a correctness criterion is local if it only requires that every transaction can be serialized along with (a subset of) the transactions committed before its last event (aborted or incomplete transactions ignored). We give a few examples of local correctness properties, including the recently proposed criterion of virtual world consistency, and present a simple though efficient implementation that satisfies non-interference and local opacity.

cs.DC↗

arXiv2013

On the Cost of Concurrency in Transactional Memory

The crux of software transactional memory (STM) is to combine an easy-to-use programming interface with an efficient utilization of the concurrent-computing abilities provided by modern machines. But does this combination come with an inherent cost? We evaluate the cost of concurrency by measuring the amount of expensive synchronization that must be employed in an STM implementation that ensures positive concurrency, i.e., allows for concurrent transaction processing in some executions. We focus on two popular progress conditions that provide positive concurrency: progressiveness and permissiveness. We show that in permissive STMs, providing a very high degree of concurrency, a transaction performs a linear number of expensive synchronization patterns with respect to its read-set size. In contrast, progressive STMs provide a very small degree of concurrency but, as we demonstrate, can be implemented using at most one expensive synchronization pattern per transaction. However, we show that even in progressive STMs, a transaction has to "protect" (e.g., by using locks or strong synchronization primitives) a linear amount of data with respect to its write-set size. Our results suggest that looking for high degrees of concurrency in STM implementations may bring a considerable synchronization cost.

cs.DC↗

arXiv2013

Safety of Deferred Update in Transactional Memory

Transactional memory allows the user to declare sequences of instructions as speculative \emph{transactions} that can either \emph{commit} or \emph{abort}. If a transaction commits, it appears to be executed sequentially, so that the committed transactions constitute a correct sequential execution. If a transaction aborts, none of its instructions can affect other transactions. The popular criterion of \emph{opacity} requires that the views of aborted transactions must also be consistent with the global sequential order constituted by committed ones. This is believed to be important, since inconsistencies observed by an aborted transaction may cause a fatal irrecoverable error or waste of the system in an infinite loop. Intuitively, an opaque implementation must ensure that no intermediate view a transaction obtains before it commits or aborts can be affected by a transaction that has not started committing yet, so called \emph{deferred-update} semantics. In this paper, we intend to grasp this intuition formally. We propose a variant of opacity that explicitly requires the sequential order to respect the deferred-update semantics. We show that our criterion is a safety property, i.e., it is prefix- and limit-closed. Unlike opacity, our property also ensures that a serialization of a history implies serializations of its prefixes. Finally, we show that our property is equivalent to opacity if we assume that no two transactions commit identical values on the same variable, and present a counter-example for scenarios when the "unique-write" assumption does not hold.

cs.DC↗