XcalableMP 2.0 and Future Directions
261
We will have a plan to design the library interface for XcalableMP programming
model, XMP API, which is aiming to provide the most equivalent programming
functions by the set of libraries.
4.4 Global Task Parallel Programming Model for Accelerators
The task-based programming recently supported in OpenMP 4.0 enables to expose
a lot of parallelism by executing several tasks of the program in the form of taskgraph. To accelerate the task-based parallel program by accelerators such as GPU
and FPGA, it is useful for some tasks frequently executed in parallel to be offloaded
to accelerators as an asynchronous task executed by accelerators.
In previous section, the global task parallel programming model is presented.
The next step will be that this global task parallel programming model is extended
to tasks offloaded to accelerators attached to each node in accelerated clusters.
Exploration of new high-performance architectures from programing model’s
point of view is an important challenge. Future parallel architecture will be more
heterogenous having many kinds of accelerators and devices attached to the nodes
and directly connected between accelerators by some dedicated interconnect. To
program such a complex and heterogenous parallel system, the global task parallel
programming model will give a flexible and decomposable model to exploit such
heterogenous high-performance architecture.
References
1. Flagship 2020 Project (Supercomputer Fugaku), https://www.r-ccs.riken.jp/en/overview/
exascalepj/
2. H. Murai, M. Sato, An efficient implementation of stencil communication for the XcalableMP
PGAS parallel programming language, in 7th International Conference on PGAS Programming Models, Edinburgh (2013)
3. FUJITSU Ltd., Development Studio uTofu User’s Guide (2020)
4. RIKEN Advanced Institute for Computational Science (RIKEN AICS), Fiber Miniapp Suite
(2104), fiber-miniapp.github.io
5. H. Murai, M. Nakao, H. Iwashita, M. Sato, Preliminary performance evaluation of coarraybased implementation of fiber Miniapp suite using XcalableMP PGAS language, in Second
Annual PGAS Applications Workshop (PAW), Denver, CO (2017)
6. CCS
QCD
Solver
benchmark
program,
https://www.ccs.tsukuba.ac.jp/qcd/
ccsqcdsolverbenchmic/
7. NTChem Overview, https://www.r-ccs.riken.jp/software_center/software/ntchem/overview/
8. K. Tsugane, J. Lee, H. Murai, M. Sato Multi-tasking execution in PGAS language XcalableMP
and communication optimization on many-core clusters, in HPC Asia 2018, Tokyo (2018), pp.
75–85
9. Joint Center for Advanced High Performance Computing (JCAHPC), Basic Specification of
Oakforest-PACS, http://jcahpc.jp/files/OFP-basic.pdf
261
We will have a plan to design the library interface for XcalableMP programming
model, XMP API, which is aiming to provide the most equivalent programming
functions by the set of libraries.
4.4 Global Task Parallel Programming Model for Accelerators
The task-based programming recently supported in OpenMP 4.0 enables to expose
a lot of parallelism by executing several tasks of the program in the form of taskgraph. To accelerate the task-based parallel program by accelerators such as GPU
and FPGA, it is useful for some tasks frequently executed in parallel to be offloaded
to accelerators as an asynchronous task executed by accelerators.
In previous section, the global task parallel programming model is presented.
The next step will be that this global task parallel programming model is extended
to tasks offloaded to accelerators attached to each node in accelerated clusters.
Exploration of new high-performance architectures from programing model’s
point of view is an important challenge. Future parallel architecture will be more
heterogenous having many kinds of accelerators and devices attached to the nodes
and directly connected between accelerators by some dedicated interconnect. To
program such a complex and heterogenous parallel system, the global task parallel
programming model will give a flexible and decomposable model to exploit such
heterogenous high-performance architecture.
References
1. Flagship 2020 Project (Supercomputer Fugaku), https://www.r-ccs.riken.jp/en/overview/
exascalepj/
2. H. Murai, M. Sato, An efficient implementation of stencil communication for the XcalableMP
PGAS parallel programming language, in 7th International Conference on PGAS Programming Models, Edinburgh (2013)
3. FUJITSU Ltd., Development Studio uTofu User’s Guide (2020)
4. RIKEN Advanced Institute for Computational Science (RIKEN AICS), Fiber Miniapp Suite
(2104), fiber-miniapp.github.io
5. H. Murai, M. Nakao, H. Iwashita, M. Sato, Preliminary performance evaluation of coarraybased implementation of fiber Miniapp suite using XcalableMP PGAS language, in Second
Annual PGAS Applications Workshop (PAW), Denver, CO (2017)
6. CCS
QCD
Solver
benchmark
program,
https://www.ccs.tsukuba.ac.jp/qcd/
ccsqcdsolverbenchmic/
7. NTChem Overview, https://www.r-ccs.riken.jp/software_center/software/ntchem/overview/
8. K. Tsugane, J. Lee, H. Murai, M. Sato Multi-tasking execution in PGAS language XcalableMP
and communication optimization on many-core clusters, in HPC Asia 2018, Tokyo (2018), pp.
75–85
9. Joint Center for Advanced High Performance Computing (JCAHPC), Basic Specification of
Oakforest-PACS, http://jcahpc.jp/files/OFP-basic.pdf