What is Adaptyst?

Introduction#

Adaptyst is an open-source and architecture-agnostic modular performance analysis tool, where modules can integrate existing work (such as profilers) and be made by external developers. The goal of the program is addressing users’ software, hardware, and system needs in a future-proof way by suggesting the most optimal compute solution for a given workflow and requirements while considering all sides of computation.

At the moment, Adaptyst concentrates on profiling and already supports all CPUs running Linux thanks to our linuxperf module which samples in detail both on-CPU and off-CPU activity of every spawned thread and process of a program, profiles low-level interactions with hardware, and more: see the linked documentation page. There’s also preliminary support of NVIDIA GPUs provided by our nvgpu module. More modules from the Adaptyst team are already planned / work in progress.

Later, through PhD research at ETH Zürich and CERN along with other potential collaborations, Adaptyst will move towards realising the goal outlined in the first paragraph, i.e. becoming a unified full-stack performance analysis and automated computer system design tool working across the entire spectrum: from embedded/edge computing to high-performance and distributed computing. This will happen while maintaining the modular design of the tool.

After going through the documentation, you are welcome to follow our blog and consult the Adaptyst management system (becoming live soon) to see the progress achieved so far towards that vision (please note that the management system is not meant to show the entire envisaged roadmap at once, so seeing few or none of roadmap-related issues at a given moment does not mean that no progress is being made).

Adaptyst flow and components#

The usage of Adaptyst starts with defining a workflow to be analysed, using one or more plugins that can be made either externally or by the Adaptyst team. This can capture a variety of scenarios from across computing, but the most common one is running a specific software code. The plugins are responsible for translating the use case(s) they represent into the Adaptyst intermediate representation in form of stateful dataflow multigraphs (SDFGs; they are part of the DaCe framework and explained here).

When the user workflow is compiled into an SDFG, it is processed by a set of system/hardware modules forming a system graph representing a computer system with nodes (like peripherals) grouped into entities (like servers): each module can be made either externally or by the Adaptyst team and it is responsible for performance analysis with respect to a specific system component (e.g. a CPU or its part, a GPU or its part, an FPGA, networking).

In the future, as the R&D work on Adaptyst progresses, the modules will take part in determining the best compute solution for the SDFG while considering all sides of computation and user requirements. This will include software optimisations, hardware choices/customisations (e.g. deciding whether a GPU is needed and what parts of the workflow should go to a GPU if so), and system adjustments. At the same time, flexibility will be maintained throughout the entire process: a user will be free to decide whether Adaptyst should concentrate more on the software, system, or hardware side when determining the compute solution. See System graphs for more details about system graphs in Adaptyst and what role they will play in automatic computer system design.

After performance analysis is finished, the results can be inspected in a variety of ways, including by our dedicated tool Adaptyst Analyser producing an interactive website. The official API for advanced result analysis is planned.

Comparison to profilers#

The table below compares Adaptyst to some of the more popular maintained profilers. It is important to note though that the tool is not meant to compete in the profiling market: it is meant to integrate existing work to provide unified full-stack performance analysis allowing both a macroscopic and microscopic view on computing performance. Additionally, Adaptyst aims to expand on that work to provide automatic optimisation suggestions while considering all sides of computation.

If there are any inaccuracies, please let us know.

¹ If supported by a user’s hardware architecture.

Comparison to software-hardware co-design and design space exploration work#

In the context of the R&D work towards realising the vision outlined in the introduction, you can have a look at the review of the existing related work here.