Scalable Tools Workshop

Autotools for Supercomputers

Christian Feld

Abstract

Developing portable software is hard. To mitigate this process, leveraging the widely adopted GNU autotools can be beneficial as they significantly reduce the effort of writing portable build systems. However, assumptions made by GNU autotools do not match the obstacles developers face on today's supercomputers with their various combinations of login- and compute-node architectures, compilers, and libraries. With Autotools for Supercomputers (AfS), we propose a modular framework on top of GNU autotools that addresses this mismatch by hiding the complexities of system diversity from developers and users, offering reasonable defaults as well as fuss-free options for customization. For developers, AfS reduces the effort of creating portable make targets to the choice for which architecture to build them (login- and/or compute-node, MPI, accelerator), while users benefit from a unified and simplified configuration process. Last but not least, build system maintainers are provided with working solutions to produce robust and maintainable systems, at the same time preventing from common pitfalls, particularly with regard to cross-compilation. AfS is actively maintained and successfully used by several projects running on a wide variety of supercomputers.

Infrastructure for Ubiquitous Performance Analysis

David Poliakoff

Abstract

In this talk we will discuss tightly integrating performance tools into applications. Building tightly-coupled performance tools enables capabilities like "always on" performance analysis, performance history tracking, and performance introspection. It also simplifies tool development by letting tools leverage application infrastructure. This talk will focus on new tool components that enable tighter application and performance tool coupling. We will discuss Gotcha, which provides a mechanism for tools to intercept application operations, and SPOT, web infrastructure for visualizing performance history. By making our tools an essential component in applications we can raise both the visibility and importance of performance tools in the HPC community.

Recent Advances in PAPI

Anthony Danalis

Abstract

PAPI has been a layer that provides a consistent interface to hardware performance counters for almost two decades. This talk will outline the newest developments and the future directions of PAPI. Energy and power, being important optimization parameters as we move to larger compute systems, have been one of the foci of PAPI development. Another recent development within PAPI has been the effort to develop means to categorize and understand the numerous counters that exist in modern hardware platforms. Lastly, this talk will introduce PAPI Software-defined events and explain our effort to extend PAPI’s role as a standardizing layer for monitoring performance of entire applications, including their interactions with external libraries, runtimes and other software layers.

Flexible Data Aggregation for Performance Profiling

David Boehme

Abstract

Many performance analysis tools in the HPC space perform some form of aggregation to compute summary information of a series of performance measurements, from summations to more complex operations like histograms. In current profilers, however, most aspects that control the aggregation are hard-coded, limiting results to certain pre-defined profile types (e.g., call-path time profiles). In this talk, we present Caliper's new aggregation approach, where users can define custom aggregation schemes in a simple SQL-like description language. This allows us to create performance profiles with application-specific data dimensions that would be difficult to obtain with traditional profiling tools.

The State and Needs of IO Tools

Greg Becker and Elsa Gonsiorowski

Abstract

I/O has historically been an afterthought in application development and performance analysis. As processing power increases and accelerators increase performance, Amdahl’s law increasingly points to I/O as an opportunity for performance optimization. Unfortunately, the tool framework for analysis of I/O performance is not as robust as that for computational performance. Tools that exist, such as IOR and Darshan, target a very small slice of I/O performance. No tool exists which measures the real-world I/O performance of an application running under real workloads, and tools that create artificial I/O to model performance do not account for network contention. Additionally complicating this space is the introduction of hierarchical storage into the I/O space. With the increased prevalence of burst buffers and other forms of hierarchical storage, we can expect to find new answers to historically well-studied problems such as optimal checkpointing frequency. As user need for I/O performance increases, there will need to be a paradigm shift in the performance tools community to including I/O in our understanding of application performance. Bringing the insights of the scalable tools community to I/O performance tools will increase our total understanding of performance in the HPC sphere.

Update on Intra- and Inter-compiler ABI Compatibility

Ben Woodard

Abstract

Another year later and libabigail’s ABI verification has improved considerably. Intra-compiler verification seems fairly solid and Inter-compiler ABI compatibility has become an area of more interest within Red Hat due to our bootstrapping of the GPGPU toolchain based on LLVM for future products and to facilitate our participation with OpenHPC. This short talk updates last year’s discussion on inter-compiler ABI compatibility. It outlines some of the continuing challenges and a few breakthroughs.

Analyzing Data Access Latency

William Jalby

Abstract

Data access latency can be a major performance issue and is very often overlooked. Data access bandwidth is considered as a major performance limiter while in some cases, the dominating factor is latency. In this talk we will present a tool for analyzing data latency impact on performance and providing some hints on data access optimization.

HPC in Ancient History

Marty Itzkowitz

Abstract

Long before there was Unix, C, C++, Java, email, internet, ..., there was still High Performance Computing. In this talk, I will describe the dominant machines of that era, the IBM 709x, and the CDC 6600 and 7600. I'll talk about the interesting characteristics of those machines, the software that ran on them, mass storage of that time, and the software development environment. I will show pictures of some of the remarkable pieces of hardware from that time, and will have a show and tell of artifacts.