Jichiang Tsai

Theoretical analysis for communication-induced checkpointing protocols with rollback-dependency trackability

Rollback-Dependency Trackability (RDT) is a property that states that all rollback dependencies between local checkpoints are on-line trackable by using a transitive dependency vector. In this paper, we address three fundamental issues in the design of communication-induced checkpointing protocols that ensure RDT. First, we prove that the following intuition commonly assumed in the literature is in fact false: If a protocol forces a checkpoint only at a stronger condition, then it must take, at most, as many forced checkpoints as a protocol based on a weaker condition. This result implies that the common approach of sharpening the checkpoint-inducing condition by piggybacking more control information on each message may not always yield a more efficient protocol. Next, we prove that there is no optimal on-line RDT protocol that takes fewer forced checkpoints than any other RDT protocol for all possible communication patterns. Finally, since comparing checkpoint-inducing conditions is not sufficient for comparing protocol performance, we present some formal techniques for comparing the performance of several existing RDT protocols.

Evaluations of domino-free communication-induced checkpointing protocols

We give a detailed analysis of communication-induced checkpointing protocols that are free of domino effect. We investigate the validity of a common intuition in the literature and demonstrate that there is no optimal on-line domino-free protocol in terms of the number of forced checkpoints. Formal proofs on comparing existing protocols in the literature are given.

Fault Tolerance for Super-Peers of P2P Systems

This paper presents an efficient fault-tolerant approach for the super-peers of peer-to-peer (P2P) file sharing systems. In the super-peer based P2P file sharing system, peers are organized into multiple groups. In each group, it has a special peer called super peer to serve the regular peers within the group. In this hierarchical architecture, if the super peer departs (fails), any file queries to its serving regular peers cannot be delivered. In the proposed approach, we propose a multiple publication technique to make each regular peer logically connect with two or more super peers in other groups. If a regular peer finds that its serving super peer cannot work, one of its other connected super peers will be selected as its new serving super peer to continuously process the file queries. To examine the effectiveness of the proposed approach, comprehensive simulations are performed to quantify the performance and overhead of the proposed approach.Several recent studies identify the memory system as the most frequent source of hardware failures in commercial servers. Techniques to protect the memory system from failures must continue to service memory requests, despite hardware failures. Furthermore, to support existing OSs, the physical address space must be retained following reconfiguration. Existing techniques either suffer from a high performance overhead or require pervasive hardware changes to support transparent recovery. In this paper, we propose physical address indirection (PAI), a lightweight, hardware-based mechanism for memory system failure recovery. PAI provides a simple hardware mapping to transparently reconstruct affected data in alternate locations, while maintaining high performance and avoiding physical address changes. With full-system simulation of commercial and scientific workloads on a 16-node distributed shared memory server, we show that prior techniques have an average degraded mode performance loss of 14 % and 51 % for commercial and scientific workloads, respectively. Using PAIs data- swap reconstruction, the same workloads have 1 % and 32 % average performance losses.

On properties of RDT communication-induced checkpointing protocols

Rollback-dependency trackability (RDT) is a property stating that all rollback dependencies between local checkpoints are online trackable by using a transitive dependency vector. The most crucial RDT characterizations introduced in the literature can be represented as certain types of RDT-PXCM-paths. Here, let the U-path and V-path be any two types of RDT-PXCM-paths. We investigate several properties of communication-induced checkpointing protocols that ensure the RDT property. First, we prove that if an online RDT protocol encounters a U-path at a point of a checkpoint and communication pattern associated with a distributed computation, it also encounters a V-path there. Moreover, if this encountered U-path is invisibly doubled, the corresponding encountered V-path is invisibly doubled as well. Therefore, we can conclude that breaking all invisibly doubled U-paths is equivalent to breaking all invisibly doubled V-paths for an online RDT protocol. Next, we continue to demonstrate that a visibly doubled U-path must contain a doubled U-cycle in the causal past. These results can further deduce that some different checkpointing protocols actually have the same behavior for all possible patterns. Finally, we present a commendatory systematic technique for comparing the performance of online RDT protocols.

On characteristics of DEF communication-induced checkpointing protocols

Domino-Effect Freedom (DEF) is a property stating that unbounded, cascading rollback propagation will not occur during the process of finding a consistent global checkpoint. DEF checkpointing protocols can be classified into two different categories: index-based and model-based. An index-based protocol timestamps local checkpoints with sequence numbers for achieving consistency, while a model-based one prohibits the formation of particular checkpoint and communication patterns in the execution. We explore several characteristics of communication-induced checkpointing protocols with the DEF property. First, we demonstrate that there is no optimal online scalar DEF protocol, in terms of the number of forced checkpoints. Then some techniques for comparing protocols are proposed. We construct some patterns or give formal proofs to compare the performance of both index-based and model-based protocols.

An efficient index-based checkpointing protocol with constant-size control information on messages

Communication-induced checkpointing (CIC) protocols can be used to prevent the domino effect. Such protocols that belong to the index-based category were shown to have a better performance. In this paper, we propose an efficient index-based CIC protocol. The fully informed (FI) protocol proposed in the literature has been known to be the best index-based CIC protocol that one can achieve since the optimal protocol needs to acquire the future information. We discover that the enhancement adopted by such a protocol rarely takes effect in practice. By discarding this enhancement, we obtain a new protocol, called NMMP. Simulation results show that our protocol is almost as efficient as FI in some typical computational environments. Especially, we demonstrate that the two protocols have the same behavior over a tree communication network. Surprisingly, NMMP only has to piggyback on each message control information of constant size, regardless of the number of processes.

On the fully-informed communication-induced checkpointing protocol

Communication-induced checkpointing (CIC) protocols can be used to prevent the domino effect. The fully-informed (FI) protocol proposed in the literature is known as the best CIC protocol so far. In this paper, we investigate some properties of such a protocol. First, we discover that an enhancement adopted by the FI protocol rarely takes effect in practice. In particular, we prove that such an enhancement is invalid over a tree communication network. Moreover, the size of the control information that the FI protocol piggybacks on application messages can be further reduced on such a type of network. Next, we show how to apply the FI protocol to another improved indexing strategy introduced in the literature. The management of a Boolean array in its control information carried on a message must be modified. Finally, we present a simulation study to analyze the properties of the FI protocol.

More Properties of Communication-Induced Checkpointing Protocols with Rollback-Dependency Trackability

Rollback-Dependency Trackability (RDT) is a property stating that all rollback dependencies between local checkpoints are on-line trackable using a transitive dependency vector. In this paper, we introduce some properties of communication-induced checkpointing protocols possessing the RDT property. First, we demonstrate that wherever an RDT protocol detects a PCM-path in the checkpoint and communication pattern associated with a distributed computation, it can also detect an EPSCM-path there. Moreover, if this detected PCM-path is non-visibly doubled, its corresponding EPSCM-path is also non-visibly doubled. Next, we go on to prove that if an RDT protocol breaks all EPSCM-cycles and non-visibly doubled EPSCM-paths, it breaks all visibly doubled EPSCM-paths as well. From these results, we find that some RDT protocols actually have the same behavior for all possible patterns. Furthermore, we also construct patterns to show that a few RDT protocols are incomparable in terms of the number of forced checkpoints. Last but not least, we discuss a simulation study to verify our previous theoretical results.

Flexible Symmetrical Global-Snapshot Algorithms for Large-Scale Distributed Systems

Most existing global-snapshot algorithms in distributed systems use control messages to coordinate the construction of a global snapshot among all processes. Since these algorithms typically assume the underlying logical overlay topology is fully connected, the number of control messages exchanged among the whole processes is proportional to the square of number of processes, resulting in higher possibility of network congestion. Hence, such algorithms are neither efficient nor scalable for a large-scale distributed system composed of a huge number of processes. Recently, some efforts have been presented to significantly reduce the number of control messages, but doing so incurs higher response time instead. In this paper, we propose an efficient global-snapshot algorithm able to let every process finish its local snapshot in a given number of rounds. Particularly, such an algorithm allows a tradeoff between the response time and the message complexity. Moreover, our global-snapshot algorithm is symmetrical in the sense that identical steps are executed by every process. This means that our algorithm is able to achieve better workload balance and less network congestion. Most importantly, based on our framework, we demonstrate that the minimum number of control messages required by a symmetrical global-snapshot algorithm is Ω(N log N), where N is the number of processes. Finally, we also assume non-FIFO channels.

Performance comparisons of index‐based communication‐induced checkpointing protocols

Abstract Communication‐induced checkpointing (CIC) protocols can be used to prevent the domino effect. Such protocols that belong to the index‐based category have been shown to perform more efficiently. In this paper, some results of comparing index‐based CIC protocols are proposed. First, we prove that comparing several protocols based on the lazy indexing strategy can be simply based on their checkpoint‐inducing conditions. Next, we show that improved indexing strategies may not always yield a better performance than the classical strategy. Finally, we present a simulation study to verify our foregoing theoretical results. The simulation is conducted in the typical point‐to‐point computational environment. Influences of enhancements on indexing strategies and checkpoint‐inducing conditions for index‐based CIC protocols are discussed.