Tuesday Schedule

Abstract:

A program's provenance encompasses the various characteristics of the
tool chain between the original source code and the resulting binary.
One obvious element of this process is the identity of the source
compiler. Recently we have developed techniques that can identify the
source compiler of binary code. Our techniques have good accuracy,
even in the case of interleaving of code from multiple compilers. In
this talk we present our current compiler inference techniques, show
how they can be used to improve the task of parsing stripped binaries,
and discuss future directions in provenance recovery.

Abstract:

The sampling rate used during performance analysis is typically chosen in an informal manner and remains fixed throughout an execution run. We have developed a mathematical framework for determining the sampling rate required to provide the best balance between achieving statistical accuracy while limiting the perturbation of the result. During execution this technique can adjust the sampling rate to adapt to changing conditions. For this talk, we present the motivation for deliberate sampling rate calculation, the straightforward mathematical framework, and the concept for the tool we are developing to implement this technique.

Abstract:

Software security vulnerabilities are an ever-present threat,
and analysts often work with programs for which only the binary
code is available.
Previous work created intrusion detection
signatures for these vulnerabilities, but required a working
exploit in advance.
In this talk, we present a method that
will generate exploit strings given only the binary code and
vulnerability location.
Our approach combines concrete and
symbolic execution techniques into a concolic execution system
for binary code, supported by a new symbolic evaluation
component for Dyninst.
We present our techniques and discuss
the expected performance of our approach.

Abstract:

Floating-point roundoff error is a well-known problem in numerical computation; it distorts results and is difficult to characterize accurately. Cancellation refers to the loss of precision as a result of subtracting nearly-identical numbers, and is a particular subset of roundoff error that can serve as a warning flag for other harder-to-detect flaws. We have developed a tool for detecting cancellation events at runtime using dynamic binary instrumentation. In this talk we will demonstrate the techniques we used and show our results on several examples.

Abstract:

High-end applications and systems are evolving towards more sophisticated modes of operation, higher levels of abstraction, and larger scales of execution. These trends challenge the state of technology for performance evaluation and optimization. While scalable toolsets for parallel and distributed system performance analysis exist, offering a range of capabilities and sophistication in performance observation and analysis, the tools are not as widely used by HPC application groups nor as broadly integrated in the application development, engineering, and tuning processes as they could and should be. This talk will discuss our recent work on simplifying performance instrumentation tasks in the TAU Performance System using DyninstAPI's binary rewriting capabilities. The talk will highlight the features of TAU and the DyninstAPI features that TAU currently uses. It will also present ideas for possible future work on improving support for binary rewriting and stack walking capabilities. A demonstration of a LiveDVD based HPC Linux distribution that includes DyninstAPI, TAU, PAPI, Scalasca and Vampir tools will be presented after the talk.