Search

We describe an architecture for nonintrusive failure detection and recovery in a cluster of Internet servers in which nodes mutually monitor their liveness and recover client sessions from failed nodes. The system is based on Backdoors, a novel architectural approach for remote healing of computer systems. Backdoors enables monitoring and recovery/repair of state in a computer system by remote access to system resources (memory, I/O devices) without using its processors. Backdoors allows remote actions to be performed with no overhead, and even when the processors (but not the memory) of a machine are not available. We have implemented a Backdoors prototype by modifying the FreeBSD kernel and using Myrinet NICs for remote access. The system uses remote DMA operations to perform monitoring, detect failures and extract OS and application state from a failed machine. We have used our system to run several open-source Internet servers and to run a complex multi-tier e-commerce application. The system tolerates multiple node failures while providing correct and continuous service to ongoing sessions, with negligible disruption.

In this paper, we propose a novel self-healing approach for a computer system based on Back-Doors (BD), a system architecture that allows monitoring and repair actions to be performed on a remote operating system or application image without using remote CPU cycles. The key ingredient of the Back-Door architecture is the Remote Memory Communication (RMC) technology provided by standards like Virtual Interface Architecture (VIA) or InfiniBand, specifically its support for remote DMA read and write operations. In this paper, we discuss the potential and challenges of the Back-Door approach to self-healing, and describe a preliminary prototype we developed as proof of concept.

We evaluate the feasibility of using Migratory TCP (M-TCP), a reliable connection-oriented transport layer protocol that supports connection migration, for building highly available Internet services. M-TCP can transparently migrate the server endpoint of a live connection and assists server applications in resuming service on migrated connections. M-TCP provides a generic solution for the problem of service continuity and availability in the face of connectivity failures. We have implemented M-TCP and present results of an experimental evaluation which shows it can efficiently provide support for highly available services. We illustrate the use of M-TCP in two applications. The first is a synthetic generic media streaming server. We show that, when the performance of the current server degrades, M-TCP an sustain throughput lose to the average server behavior by migrating connections to better servers. The second application is a transactional database server in which we have integrated support for migrating client connections. Using our system, a database frontend an continue the execution of a series of transactions submitted by a remote client in a session started with another front-end. The system allows a session to survive adverse conditions by connection migration, while ensuring that ACID semantics are preserved and that the execution is deterministic across migration.

This paper presents a fault tolerance algorithm for a home-based lazy release consistency distributed shared memory (DSM) system based on volatile logging and independent checkpointing. The proposed approach targets large-scale distributed shared-memory computing on local-area clusters of computers as well as collaborative shared-memory applications on wide-area meta-clusters over the Internet. The challenge in building such systems lies in controlling the size of the logs and to garbage collect the unnecessary checkpoints in the absence of global coordination. In this paper we define a set of rules for lazy log trimming (LLT) and checkpoint garbage collection (CGC) and prove that they do not affect the recoverability of the system. We have implemented our logging algorithm in a home-based DSM system and showed on three representative applications that our scheme effectively bounds the size of the logs and the number of checkpointed page versions kept in stable storage.