Scalable Tools Workshop

I/O Recording and Workflow Analysis with Score-P and Vampir

Bill Williams

Abstract

As HPC tasks become larger and more complex, developers attempt to address this complexity by decomposing problems into a workflow of simpler applications, rather than a single application that can handle everything. These workflows present multiple challenges to the tools community, but one of the chief problems is the analysis of communication between jobs. In many workflows, the filesystem plays an analogous role in inter-job communication to that of MPI in interprocess communication. This suggests that we should adapt familiar techniques from MPI instrumentation and performance analysis to the I/O domain. We present initial work in enabling this analysis: * New I/O recording features in the ScoreP measurement system * A prototype suite of tools for HPC workflow instrumentation * A prototype visualisation tool for analysing and displaying the results of this instrumentation * Preliminary results in automatically determining the I/O dependencies within a workflow Our ultimate goal is to provide the user with information about how they can restructure their workflow in order to increase parallelism, improve the efficiency of their use of HPC resources, and reduce or eliminate false dependencies between jobs.

Exploiting Modern Hardware Features via Lightweight Profiling

Probir Roy

Abstract

Over the past few decades, computer systems have evolved into two directions: more parallelism and deeper memory hierarchies. Software needs to make an adaption to efficiently map itself to the modern and emerging architectures for the optimal performance. My research involves building various tools to pinpoint inefficient hardware- software interactions and guide code optimization. In this talk, I will discuss different forms of inefficiencies originating from modern computer architectures: contention across simultaneous multithreading, cache conflicts, and bandwidth under-utilization. I will then present novel analysis tools to identify these inefficiencies with low overhead. Guided by our tools, we can successfully optimize program inefficiencies, which yields significant speedups.

Ubiquitous Performance Analysis

David Boehme

Abstract

Performance Analysis of Very Large Shared Memory Systems Using Intel(R) Optane DC Memory

Ramesh Peri

Abstract

There is a new kind of memory that was announced recently called Intel(R) Optane DC memory which is lot more dense than traditional DRAM but with more access latency. This introduces a new layer in the memory hierarchy and allows systems with many TB of main memory to be built and use existing DDR4 as another layer of cache. In this talk I will introduce the tools we built to analyze performance on such large scale systems. I will also share our experiences in using the performance tools to optimize graph analytics algorithms on very large graphs, containing billions of vertices and trillions of edges, on this system.

MuMMI: Exploring Performance and Power Trade-offs

Xingfu Wu, Valerie Taylor

Abstract

Interfaces for Runtime Correctness Checking of Parallel Programs

Joachim Protze

Abstract