Check out the new USENIX Web site.
SysML08 Banner

OSDI '08


Thursday, December 11, 2008

Reinforcement Learning for Optimizing Flash Memory Cache Management
Byung Kon Kang, Kee-Eung Kim, Wook Jung, and Jin-Soo Kim, KAIST

The flash memory is rapidly becoming a popular storage medium as portable electronic devices grow prevalent. However, one shortcoming of the flash memory is the asymmetry between its read and write costs. This characteristic becomes crucial for designing cache management algorithms for flash memory, because evicting a dirty page from the cache will lead to a costly write operation. Traditional cache management algorithms fail to take this characteristic into account, since they are designed to optimize the hit rate only, rather than the overall flash memory operation cost. In this paper, we present a reinforcement learning approach to the cache management for flash memory that addresses the overall access cost, borrowing some of the insights from a recently proposed heuristic algorithm. We experiment on the trace of the real application usage data, and show that the systematically optimized cache management algorithm outperforms the heuristic algorithm.

Predictive System Health Monitoring for Large-Scale Hosting Infrastructures
Xiaohui Gu, North Carolina State University; Haixun Wang, IBM T.J. Watson Research

Large-scale hosting infrastructures have become important platforms for many real-world systems such as data stream processing, virtual computing lab, and enterprise data centers. However, system administrators are often overwhelmed by the tasks of correcting various system health problems such as processing bottlenecks and service level objective (SLO) violations. In this paper, we present ALERT, a 24x7 predictive health monitoring system for large-scale hosting infrastructures. ALERT employs light-weight stream learning methods to achieve online anomaly prediction. Particularly, ALERT achieves {\em robust} anomaly prediction for dynamic systems using self-evolving stream classification methods. We have implemented a prototype of the ALERT system and tested it on the IBM System S stream processing cluster. Our experiments show that ALERT can achieve much higher prediction accuracy than conventional learning methods while imposing low overhead to the hosting infrastructure.

Ganesha: Black-Box Fault Diagnosis for MapReduce Systems
Xinghao Pan, Jiaqi Tan, Soila Kavulya, Rajeev Gandhi, and Priya Narasimhan, Carnegie Mellon University

Ganesha aims to diagnose faults transparently in MapReduce systems, by analyzing OS-level metrics alone. Ganesha's approach is based on peer-symmetry under fault-free conditions, and can diagnose faults that manifest asymmetrically at nodes within a MapReduce system. While our training is performed on smaller Hadoop clusters and for specific workloads, our approach allows us to diagnose faults in larger Hadoop clusters and for unencountered workloads. We also candidly highlight faults that escape Ganesha's black-box diagnosis.

Using Machine Learning to Auto-tune a Stencil Code on a Multicore Architecture
Archana Ganapathi, Kaushik Datta, Armando Fox, and David Patterson, University of California, Berkeley

The computing industry is undergoing a multicore revolution. As a result, architectures have become so complex and diverse that the compiler alone is unable to fully utilize system resources. Auto-tuning has emerged as an effective method to get good performance across different platforms. However, the parameter space generated by autotuning is often intractable and may take weeks to explore. We propose a technique that uses machine learning to guide auto-tuning. Our technique produces performance within 0.5% of human expert-optimized tuning.

Generalization of Machine Learning Based Encrypted Traffic Identification: A Comparison of Two Classifiers
Riyad Alshammari and Nur Zincir-Heywood, Dalhousie University

The objective of this work is to assess the generalization of machine learning based traffic identification for classifying encrypted traffic where SSH is taken as an example application. Here what we mean by generalization is that the classifiers are trained on data from one network but tested on data from an entirely different network. To this end, two learning algorithms -- RIPPER and C4.5 -- are evaluated using flow based features, where IP addresses, source/destination ports and payload information are not employed. Results of 10 runs for each learning algorithm indicate that a high detection rate and a low false positive rate can be achieved using C4.5 or RIPPER. However, when these algorithms are tested for their generalization by training on a data from one network and testing on data from another network, C4.5 algorithm performs much better than RIPPER algorithm.

Kudzu: A Self-Balancing P2P File Transfer System
Sean Barker, Marius Catalin Iordan, and Jeannie Albrecht, Williams College; Barath Raghavan, University of California, San Diego

Many peer to peer file sharing systems to date rely in part upon centralized components or rigid network structure. The former approach typically sacrifices robustness for efficiency whereas the latter sacrifices efficiency for robustness. In this paper we present Kudzu, a system that is both fully decentralized and unstructured, yet outperforms today's most popular file sharing system, BitTorrent. Key to Kudzu's design is its automatic peer structure rebalancing, which obviates the usual explicit peer management protocols used by peer to peer systems. We examine the design of our network architecture and discuss its relationship to existing file transfer systems. We also provide an empirical evaluation of our current implementation's performance on a series of tests conducted on the PlanetLab testbed. We evaluate download speeds, load capabilities, and bandwidth efficiency.

? Need help? Use our Contacts page.

Last changed: 26 Nov. 2008 jel