I-RMI: Performance Isolation in Information Flow Applications

Abstract

Abstract. A problem with many distributed applications is their behavior in lieu of unpredictable variations in user request volumes or in available resources. This paper explores a performance isolation-based approach to creating robust distributed applications. For each application, the approach is to (1) understand the performance dependencies that pervade it and then (2) provide mechanisms for imposing constraints on the possible 'spread' of such dependencies through the application. Concrete results are attained for J2EE middleware, for which we identify sample performance dependencies: in the application layer during request execution and in the middleware layer during request de-fragmentation and during return parameter marshalling. Isolation points are the novel software abstraction used to capture performance dependencies and represent solutions for dealing with them, and they are used to create (2) I(solation)-RMI, which is a version of RMI-IIOP implemented in the WebSphere service infrastructure enhanced with isolation points. Initial results show the approach's ability to detect and filter ill-behaving messages that can cause an up to a 85% drop in performance for the Trade3 benchmark, and to eliminate up to a 56% drop in performance due to misbehaving clients.

Utility-Driven Proactive Management of Availability in Enterprise-Scale Information Flows

“…For example, one of our collaborators, reported an occasional surge in the usage of resources connected to their Operational Information System (OIS) [15] that traced back to a particular uncommon message type. The resulting performance hit caused other subsystem's requests to build up, including those from the front ends used by clients, ultimately threatening operational failure (e.g., inappropriately long response times) or revenue loss (e.g., clients going to alternate sites).…”

I-Queue: Smart Queues for Service Management

“…Platform virtualization [1]and automated resource monitoring and management [2,3] are system-level contributions to this domain. Middleware developers have introduced new functionality like automated configuration management [4],improved operator interfaces like Tivoli's 'dashboards' [5],automated methods for performance understanding and display [6], and new methods for limiting the potential effects of failures [7,8]. Large efforts like IBM's Autonomic Computing and HP's Adaptive Enterprise initiativesare developing ways to automate complex management or configuration tasks,creating new management standards ranging from Common Base Events for representing monitoring information [3] to means for stating application-level policies or component requirements (e.g.…”

“…In earlier work, we identified and found ways to deal with a simple case observed by one of our industry partners, which concerned single requests, termed a 'poison message' that consistently caused unusual system and application responses [8]. In this paper, we tackle the more complex problem of sets or sequences of requests that cause such behaviors, and where such problems may depend on dynamic system conditions, such as which business rules are currently being run or to what current states they are being applied.…”

“…These techniques use dynamic pattern matching to detect poison sequences. While failure detection uses general methods, the techniques we use for failure prevention exploit application semantics, similar to what we have done in our earlier work on poison messages and more generally, in our 'Isolation-RMI' implementation of an improved communication infrastructure for Java-based enterprise infrastructures like Websphere or JBOSS [8]. As with solutions used to improve the performance of 3-tier web service infrastructures [14], we simply interpose a request scheduler between clients and server.…”

“…The I-Queue solution goes beyond addressing the specific server-based problemoutlined above, for multiple reasons. First, I-Queue is another element of themore general solution for performance and behavior isolation for distributedenterprise applications described in [8]. The basic idea of that solutionis to embed performance monitoring and associated management functionality intokey interfaces of modern enterprise middleware: (1) component interfaces, (2) communication substrates like RMI, and (3) middleware-system interfaces.…”