HAM - Heterogeneous Active Messages for Offloading

	master	devel
Build
Coverity		NA

HAM (Heterogeneous Active Messages) is a research project on efficient active messaging between heterogeneous architectures and binary files. Its original motivation is to efficiently offload computations onto accelerators like the Intel Xeon Phi - locally as well as remotely over a fabric.

The design goals are:

a simple API unifying intra- and internode offloading,
minimising the overhead cost of offloading a computation to another compute device, and
to implement offloading in pure C++, i.e. not introducing yet another language/compiler extension, or tool dependencies

The outcome of this project are the two libraries in this repository:

HAM, a Heterogeneous Active Messages C++ library providing RPC (Remote Procedure Call) functionality based on template meta-programming
HAM-Offload, an implementation of the offload programming model based on HAM with support for a wide variety of host/target hardware, including x86-64, Intel Xeon Phi (KNC/KNL), NEC SX-Aurora TSUBASA, and ARM64

In general, HAM-Offload can be used to create multi-process applications, where one process is a logical host running the program's main, and the other processes being offload targets (asynchronously) executing computations, i.e. function calls, that were offloaded by the host process. The API also provides functionality to move data between the processes. The participating processes communicate through one of the supported communication backends.

These capabilites can be combined into different application patterns, like local offloading, offload over fabric, offloading to another (remote) CPU instead of an accelerator, offloading from an x86 host to a remote ARM64 node or vice versa, or reverse offloading from an accelerator running the host process onto one or multiple CPUs running the target processes - be creative.

See the Wiki for further documentation.

Feedback like bug reports, feature requests, or performance results are always welcome. Please create an issue or send an e-mail.

Dependencies

	requirement	version
Compiler	Standard C++ compiler	≥ C++11
Build System	CMake	≥ 3.2
Libraries	At least one supported communication library:
	- MPI (MPICH, OpenMPI, Intel MPI, NEC MPI, ...)	≥ MPI-2
	- for Intel Xeon Phi (KNC): MPSS with SCIF (or MPI)	any
	- for NEC SX-Aurora TSUBASA, Vector Engine: NEC VEOS+VEO	≥ 2.x
	Everything else is header-only and shipped in ham/thirdparty

Quickstart

NOTE: we assume some MPI implementation to be available here.

Download:

git clone https://github.com/noma/ham.git
cd ham

Build:

mkdir build
cd build
cmake ../ham
make -j

Run:

mpirun -n 2 ./inner_product_mpi

Examples

A simple example that demonstrates the usage of HAM-Offload can be found in: ham/src/inner_product_cpp.

An example CMake project using HAM-Offload can be found in: examples/hello_world.

Version History

version	notes
v0.3	- added support for the NEC SX-Aurora TSUBASA
	- replaced Boost Program.Options with CLI11 (header-only)
v0.2	- minor fixes
v0.1	- initial release with Xeon Phi (KNC) support via SCIF and MPI

Publications and Citing

M. Noack, Heterogeneous Active Messages (HAM): Implementing Lightweight Remote Procedure Calls in C++, IWOCL/DHPCC++'19, DOI: 10.1145/3318170.3318195
M. Noack, E. Focht, Th. Steinke, Heterogeneous Active Messages for Offloading on the NEC SX-Aurora TSUBASA IPDPS/HCW'19, DOI: IPDPSW.2019.00014
M. Noack, F. Wende, Th. Steinke, F. Cordes, A Unified Programming Model for Intra- and Inter-Node Offloading on Xeon Phi Clusters, SC'14, DOI: 10.1109/SC.2014.22

Acknowledgements

This work is supported by the ZIB (Zuse Institute Berlin) and the IPCC (Intel Parallel Computing Center) activities there.

The architecture of the system was influenced by earlier work with and on the TACO framework.

noma / ham