BigPicture

The Big Picture

Introduction

The AllScale (http://www.allscale.eu) environment aims at facilitating the development of parallel application codes for very large distributed memory systems without application developers having to concern themselves with details of the underlying runtime systems and hardware.

The following figure shows a topological overview of the components of AllScale:

Responsibilities

Component	Developed and maintained by	Repository
API	University of Innsbruck, Austria	https://github.com/allscale/allscale_api
Compiler	University of Innsbruck, Austria	https://github.com/allscale/allscale_compiler
Runtime	University of Erlangen-Nuremberg, Germany	https://github.com/allscale/allscale_runtime
Monitoring	KTH Royal Institute of Technology, Sweden	https://github.com/allscale/allscale_runtime
Resilience	Queen's University Belfast, Northern Ireland	https://github.com/allscale/allscale_runtime
Scheduler	IBM Research, Ireland	https://github.com/allscale/allscale_runtime

Toolchains

The AllScale environment combines two essential aspects in developing large scale parallel applications. First, it aims at maximizing the developer's productivity during application development, requiring a development setup with a small footprint and few dependencies. Second, it needs to offer scalable performance and resilience for large distributed memory systems, requiring additional tools such as a resilience manager.

In order to meet both requirements, the AllScale environment provides two toolchains, one for each use case:

• The Standard toolchain supports the development of applications using the AllScale API on shared memory systems with a minimal footprint, only requiring a C++14-compliant compiler and standard library.
• The AllScale toolchain supports the deployment of applications using the AllScale API on large distributed memory systems, featuring resilience, multi-objective optimization and developer feedback.

Toolchain Features and Use Cases

	Standard Toolchain	AllScale Toolchain
Supports parallel operator primitives and data structures	✓	✓
Requires C++14-compliant compiler	✓	✓
Supports shared memory parallelism	✓	✓
Supports short development cycles	✓
Supports standard development and debugging tools	✓
Requires third party libraries		✓
Supports distributed memory parallelism		✓
Supports resilience features		✓
Supports OpenCL devices		✓
Supports multi-objective optimization		✓
Supports online monitoring		✓