Yu Chen

BitVault: a highly reliable distributed data retention platform

This paper summarizes our experience designing and implementing BitVault: a content-addressable retention platform for large volumes of reference data -- seldom-changing information that needs to be retained for a long time. BitVault uses smart bricks as the building block to lower the hardware cost. The challenges are to keep management costs low in a system that scales from one brick to tens of thousands, to ensure reliability, and to deliver a simple design. Our design incorporates peer-to-peer (P2P) technologies for self-managing and self-healing and uses massively parallel repair to reduce system vulnerability to data loss. The simplicity of the architecture relies on an eventually reliable membership service provided by a perfect one-hop distributed hash table (DHT). Its object-driven repair model yields last-replica recall guarantee independent of the failure scenario. So long as the last copy of a data object remains in the system, that data can be retrieved and its replication degree can be restored. A prototype has been implemented. Theoretical analysis, simulations and experiments have been conducted to validate the design of BitVault.

Weakening failure detectors for k-set agreement via the partition approach

In this paper, we propose the partition approach and define several new classes of partitioned failure detectors weaker than existing failure detectors for the k-set agreement problem in both the shared-memory model and the message-passing model. In the shared-memory model with n + 1 processes, for any 2 ≤ k ≤ n, we first propose a partitioned failure detector ΠΩk that solves k-set agreement with shared read/write registers and is strictly weaker than Ωk, which was conjectured to be the weakest failure detector for k-set agreement in the shared-memory model [19]. We then propose a series of partitioned failure detectors that can solve n-set agreement, yet they are strictly weaker than γ [10], the weakest failure detector ever found before our work to circumvent any asynchronous impossible problems in the shared-memory model. We also define two new families of partitioned failure detectors in the message-passing model that are strictly weaker than the existing ones for k-set agreement. Our results demonstrate that the partition approach opens a new dimension for weakening failure detectors related to set agreement, and it is an effective approach to check whether a failure detector is the weakest one or not for set agreement. So far, all previous candidates for the weakest failure detectors of set agreement have been disproved by the partitioned failure detectors.

Failure Detectors and Extended Paxos for k-Set Agreement

Soft errors resulting from the impact of charged particles are emerging as a major issue in the design of reliable circuits at deep sub-micron dimensions. In this paper, we model the sensitivity of individual circuit classes to single event upsets using predictive technology models over a range of CMOS device sizes from 90 nm down to 32 nm. Modeling the relative position of particle strikes as injected current pulses of varying amplitude and fall time, we find that the critical charge for each technology is an almost linear function both of the fall time of the injected current and the supply voltage. This simple relationship will simplify the task of estimating circuit-level soft error rate (SER) and support the development of an efficient SER modeling and optimization tool that might eventually be integrated into a high level language design flow.Failure detector class Omegakappa has been defined in (G. Neiger, 1995) as an extension to failure detector Omega, and an algorithm has been given in (A. Mostefaoui et al., 2005) to solve k-set agreement using Omegakappa in asynchronous message-passing systems. In this paper, we extend these previous work in two directions. First, we define two new classes of failure detectors Omegakappa and Omegakappa ,which are new ways of extending Omega and show that they are equivalent to Omegakappa. Class Omegakappa is more flexible than Omegakappa in that it does not require the outputs to stabilize eventually, while class Omegakappa does not refer to other processes in its outputs. Second, we present a new algorithm that solves k-set agreement using Omegakappa when a majority of processes do not crash. The algorithm is a faithful extension of the Paxos algorithm (L. Lamport, 1998), and thus it inherits the efficiency, flexibility, and robustness of the Paxos algorithm. In particular, it has better message complexity than the algorithm in (A. Mostefaoui et al., 2005). Both the new failure detectors and the new algorithm enrich our understanding of the k-set agreement problem.

Partition approach to failure detectors for k-set agreement

In k-set agreement problem, every process proposes a value and eventually at most k different values can be decided. When k > 1, different subset of processes may decide on different values, and thus it naturally exhibits partition among processes based on their decision values. In this paper, we propose the partition approach to dene failure detectors that capture the partition nature of k-set agreement. The power of the partition approach is to further weaken failure detectors that are already very weak in solving k-set agreement, and thus invalid the failure detectors as candidates for the weakest failure detectors for k-set agreement. Using the approach, we propose two new classes of failure detectors, statically partitioned failure detectors k and splittable partitioned failure detectors S k , both are strong enough to solve k-set agreement in the message passing model. However, we show that k is strictly weaker than k, the weakest failure detectors known so far for k-set agreement, and S is even weaker than k. The partition approach provides a new dimension to weaken failure detectors related to k-set agreement. It is an effective way to check whether a failure detector is the weakest one solving k-set agreement or not. Together with [4], we show that so far all candidates for the weakest failure detectors including k and in both the message-passing model and the shared-memory model have failed our partition test.

Decentralized, connectivity-preserving, and cost-effective structured overlay maintenance

In this paper we present a rigorous treatment to structured overlay maintenance in decentralized peer-to-peer (P2P) systems subject to various system and network failures. We present a precise specification that requires the overlay maintenance protocols to be decentralized, preserve overlay connectivity, always converge to the desired structure whenever possible, and only maintain a small local state independent of the size of the system. We then provide a complete protocol with proof showing that it satisfies the specification. The protocol solves a number of subtle issues caused by decentralization and concurrency in the system. Our specification and the protocol overcomes a number of limitations of existing overlay maintenance protocols, such as the reliance on a centralized and continuously available bootstrap system, the assumption of a known system stabilization time, and the need to maintain large local membership lists.

Brief announcement: decentralized, connectivity-preserving, and cost-effective structured overlay maintenance

Since their introduction, structured overlays have been used as an important substrate for many peer-to-peer applications. In a structured peer-to-peer overlay, each node maintains a partial list of other nodes in the system, and these partial lists together form an overlay topology that satisfies certain structural properties (e.g., a ring). Various system conditions, such as node joins and leaves, message delays and network partitions, affect overlay topology, so overlay topology should adjust itself appropriately to maintain structural properties. Topology maintenance is crucial to the correctness and the performance of applications built on top of the overlay.