[lk] livio soares | research

[Peter Teoh]

http://www.eecg.toronto.edu/~livio/research.shtml My research is currently focused on the performance of systems software. More specifically, I am investigating the interactions between modern computer architectures and systems software. One useful tool which I have worked with is the hardware performance countersfound in most modern processors. I investigate software scalability in the face of current chip-multiprocessor and multithreaded hardware trends. My research at the University of Toronto has been performanced under the guidance of Michael Stumm. A few years ago, I had the opportunity to hack the K42 research operating system, from which I have learned a great deal about operating-system structure, design and implementation. K42 is a research operating system designed for shared-memory multiprocessors with scalability and flexibility as its main goals. For my Masters, I developed the K42 File System (KFS), including a concurrent port to the Linux kernel. The goal of KFS was to build an infra-structure for a flexible file-system where each file or directory can have distinct implementations (layout on disk and caching policies). During KFS' development I also developed a novel file-system consistency technique called meta-data snapshotting, which was possible given KFS' unique decentralized meta-data structure. This work was done at the University of São Paulo, under the guidance ofDilma Da Silva.   Paper Trail (a.k.a, Publications) 2011 Exception-less System Calls for Event-Driven Servers Livio Soares and Michael Stumm  USENIX Annual Technical Conference (USENIX ATC'11)  To appear. Mind the Gap: Reconnecting Architecture and OS Research Jeffrey C. Mogul, Andrew Baumman, Timothy Roscoe, and Livio Soares  13th Workshop on Hot Topics in Operating Systems (HotOS 2011)  To appear. 2010 FlexSC: Flexible System Call Scheduling with Exception-Less System Calls Slides: PDF [5.1MB]  Presentation video: hosted at USENIX, may require membership  Livio Soares and Michael Stumm  9th USENIX Symposium on Operating Systems Design and Implementation (OSDI 2010)  This work questions the 30+ year old system call mechanism operating systems use by quantifying the performace impact, at the processor level, of system calls for server workloads. We propose a new exception-less mechanism for system calls. Excepion-less system calls allows for flexibly scheduling operating system work, enabling run-time systems to greatly improve locality of execution, and consequently, performance. One interesting use-case of exception-less system calls is dynamic core specialization, where system calls are dynamically redirected to specific cores in a multi-core system, adapting to workload needs. We also describe a new threading package that transparently translates legacy system calls into exception-less ones. Using a Linux implementation of exception-less system calls (FlexSC) and a pthreads-compatible thread package (called FlexSC-Threads), we demonstrate a 2x speed-up for internet servers such as Apache and BIND. 2009 RapidMRC: Approximating L2 Miss Rate Curves on Commodity Systems for Online Optimizations David Tam, Reza Azimi, Livio Soares, and Michael Stumm  In the 14th International Conference on Architectural Support for Programming Languages and Operating Systems (ASPLOS 2009)  RapidMRC is an online tool for generating L2 MRCs targeted at dynamic optimizations in the run-time system. This work describes the underlying technique using Power5 hardware performance counters, as well as an interesting use-case for L2 MRCs: cache partitioning in a multi-core system. Enhancing Operating System Support for Multicore Processors by Using Hardware Performance Monitoring Reza Azimi, David Tam, Livio Soares, and Michael Stumm  ACM SIGOPS Operating System Review (OSR) - Special issue on The Interaction Among the OS, the Compiler, and Multicore Processors, April 2009  This paper describes the work our group has been pursuing to improve performance of multi-core systems. We focused on dynamic operating system techniques that adapt using run-time hardware performance counters information. 2008 Reducing the Harmful Effects of Last-Level Cache Polluters with an OS-Level, Software-Only Pollute Buffer Slides: PPT, PDF  Livio Soares, David Tam and Michael Stumm  In the 41st International Symposium on Microarchitecture (MICRO-41), Lake Como, Italy, 2008  One of the eight papers nominated for the Best Paper Award by the Program Committee.  Describes a new software mechanism, "pollute buffer", that improves processor cache performance of cache/memory intensive applications. The OS profiles application's address spaces with hardware performance counters and re-maps polluter pages to a small partition of the processor cache. Performance of 5% to 35% improvement is obtained on commodity hardware. 2007 PATH: Page Access Tracking to Improve Memory Management Reza Azimi, Livio Soares, Michael Stumm, Angela Demke-Brown, and Tom Walsh  In International Symposium on Memory Management (ISMM 2007) Experience Distributing Objects in an SMMP OS Jonathan Appavoo, Dilma Da Silva, Orran Krieger, Marc Auslander, Michal Ostrowski, Bryan Rosenburg, Amos Waterland, Robert W. Wisniewski, Jimi Xenidis, Michael Stumm, Livio Soares  In ACM Transactions on Computer Systems (TOCS) [2007, Volume 25, Issue 3] Managing Shared L2 Caches on Multicore Systems in Software David Tam, Reza Azimi, Livio Soares, and Michael Stumm  Workshop on the Interaction between Operating Systems and Computer Architecture (WIOSCA 2007) [Held with ISCA-2007] Experiences Understanding Performance in a Commercial Scale-Out Environment Robert W. Wisniewski, Reza Azimi, Mathieu Desnoyers, Maged M. Michael, Jose Moreira, Doron Shiloach, and Livio Soares  In European Conference on Parallel Computing (Euro-Par 2007) 2004 KFS: Exploring Flexibility in File System Design Dilma Da Silva, Livio Soares, and Orran Krieger  In the Brazilian Workshop on Operating Systems (WSO'2004)  It is also available on the K42 web page here (Poster, ps.gz) Meta-data Snapshotting: The Quest for File System Consistency Livio Soares, Orran Krieger, Dilma Da Silva  In the Poster Session of the 19th SOSP 2003 Meta-data Snapshotting: A Simple Mechanism for File System Consistency Livio Soares, Orran Krieger, Dilma Da Silva  In International Workshop on Storage Network Architecture and Parallel I/O (SNAPI '03) [Held with PACT-2003]  It is also available on the K42 web page here, and on the SNAPI pageA here.  We propose meta-data snapshotting, a new, low overhead, technique for maintaining file-system meta-data consistency. The main idea consists of continually snapshotting the meta-data hierarchy through "copy-on-write". The technique is considerably simpler than soft-updates, and performs better than journalling. The paper describes our Linux implementation using the K42 file system, and performance comparisons with "ext3" journalled file system. Master's Thesis (in Portuguese): Flexibilidade em um Sistema de Arquivos Componentizado (Flexibility in a Component-based File System) Livio Soares  Thesis presented to the University of São Paulo as requirement for my Master's of Science in Computer Science  Advisor: Dilma Da Silva.