Incremental Parallel and Distributed Systems

Pramod Bhatotia MPI-SWS Monday, May 5, 2014 @ 10:00 am Room BC 420 Hosted by: Prof. George Candea

Abstract

Many applications exhibit the property that small, localized changes to the input result in only minor changes to the output. Incremental computations seek to exploit this observation to speed-up successive runs of an application by tracking control and data dependencies, and re-executing only those parts of the computation that are affected by a given input change. To realize the benefits of incremental computation, researchers and practitioners are developing new systems where the programmer can provide efficient update mechanisms for changing application data. Unfortunately, most of the existing solutions are limiting because they either depart from existing programming models, or require programmers to devise an incremental update mechanism (or a dynamic algorithm) on a per-application basis. The high-level goal of my thesis research is to enable practical, automatic, and efficient incremental computations in parallel and distributed systems. My approach neither requires a radical departure from current models of programming nor complex, application-specific dynamic algorithms. In this talk, I will first present a high-level description of my thesis research. Thereafter, as a concrete example I will describe my recent project “iThreads”. iThreads is a POSIX-compliant threading library to support parallel incremental computation targeting unmodified C/C++ multithreaded programs. The iThreads library can be used as a drop-in replacement for the pthreads library, making it trivial to obtain the benefits of incremental computation by a simple exchange of libraries linked, without even recompiling the application code. To achieve this result, we design our algorithms and implementation to operate at the compiled binary code level by leveraging operating system-specific mechanisms. Our design choice of iThreads tries to strike a balance between efficiency of incremental computation and transparency for unmodified C/C++ pthreads-based programs. Our evaluation on a multicore platform using benchmarks from the PARSEC and Phoenix suites shows significant performance improvements for the majority of applications.

Biography

I’m a PhD student in the Sysnets group at MPI-SWS. My advisors are Rodrigo Rodrigues and Umut Acar. Prior to joining MPI-SWS, I worked as a member of technical staff at Adobe Systems (till the end of 2009) and in the HPC group (2008-2009) at IBM Research–India. Before that, I was a graduate student at IIT Kanpur. I spent the summer of 2012 as a research intern in the Systems and Networking group at Microsoft Research Cambridge, working with Dushyanth Narayanan.