246
M. Sato et al.
The XcalableMP project has been started from 2008 and the discussion on
XcalableMP 1.x has converged. We are now working on a new version, XcalableMP
2.0, targeted for cutting-edge high-performance systems with manycore processors
by multithreading and multi-tasking with integrations of PGAS model and synchronization models. In this new programming model, the execution of the program is
decomposed into several tasks executed according the dependency between tasks.
This model will enable less overhead of synchronization by eliminating expensive
global synchronization, overlap between computation and communication in manycore, and light-weight communication by RDMA in PGAS model. We will extend
this programming model to combine several kinds of accelerators such as GPU,
FPGA, and special-purpose processors with large-scale general-purpose manycore
systems. It enables some tasks to be offloaded into the accelerators such as FPGA
as well as each core in modern manycore processor. We consider this configuration
as a general global architecture of the future system as some part of system will
be specialized for high performance and power efficiency. Our programming model
will make it easy to adopt the existing computational science program to the new
systems.
In Sect. 3, a proposal for XcalableMP 2.0 is presented, followed by retrospectives
and challenges for future PGAS models in Sect. 4.
2 XcalableMP on Fugaku
In this section, we report our early experience and the preliminary performance of
XcalableMP on Fugaku. The Fugaku is a huge-scale system with general-purpose
manycore processors. The node processor is a single chip, Fujitsu A64FX, which
consists of 48 cores with 2 or 4 cores dedicated for OS activities, 32 GiB HBM2
memory, with Tofu-D interconnect, and a PCI express controller in the chip together.
The Fugaku system consists of 158,976 nodes in 432 racks. The Fugaku is scheduled
to be put into operation for public service around 2021. In 2020, the installation is
completed and the system partially serves the early access program.
XcalableMP is available as a parallel programming language for the Fugaku,
supported by R-CCS team with Fujitsu. C and Fortran are supported as base
languages with XcalableMP 1.2 compliant.
We report the preliminary performance of XcalableMP program running on the
Fugaku. 1
We used the following versions:
• Omni XcalableMP Version: 1.3.2, Git Hash: 6d23f46.
• Language specification: 1.2.25.
1 The reported results were obtained on the evaluation environment in the trial phase. Note that the
performance is not guaranteed at the start of its operation.
M. Sato et al.
The XcalableMP project has been started from 2008 and the discussion on
XcalableMP 1.x has converged. We are now working on a new version, XcalableMP
2.0, targeted for cutting-edge high-performance systems with manycore processors
by multithreading and multi-tasking with integrations of PGAS model and synchronization models. In this new programming model, the execution of the program is
decomposed into several tasks executed according the dependency between tasks.
This model will enable less overhead of synchronization by eliminating expensive
global synchronization, overlap between computation and communication in manycore, and light-weight communication by RDMA in PGAS model. We will extend
this programming model to combine several kinds of accelerators such as GPU,
FPGA, and special-purpose processors with large-scale general-purpose manycore
systems. It enables some tasks to be offloaded into the accelerators such as FPGA
as well as each core in modern manycore processor. We consider this configuration
as a general global architecture of the future system as some part of system will
be specialized for high performance and power efficiency. Our programming model
will make it easy to adopt the existing computational science program to the new
systems.
In Sect. 3, a proposal for XcalableMP 2.0 is presented, followed by retrospectives
and challenges for future PGAS models in Sect. 4.
2 XcalableMP on Fugaku
In this section, we report our early experience and the preliminary performance of
XcalableMP on Fugaku. The Fugaku is a huge-scale system with general-purpose
manycore processors. The node processor is a single chip, Fujitsu A64FX, which
consists of 48 cores with 2 or 4 cores dedicated for OS activities, 32 GiB HBM2
memory, with Tofu-D interconnect, and a PCI express controller in the chip together.
The Fugaku system consists of 158,976 nodes in 432 racks. The Fugaku is scheduled
to be put into operation for public service around 2021. In 2020, the installation is
completed and the system partially serves the early access program.
XcalableMP is available as a parallel programming language for the Fugaku,
supported by R-CCS team with Fujitsu. C and Fortran are supported as base
languages with XcalableMP 1.2 compliant.
We report the preliminary performance of XcalableMP program running on the
Fugaku. 1
We used the following versions:
• Omni XcalableMP Version: 1.3.2, Git Hash: 6d23f46.
• Language specification: 1.2.25.
1 The reported results were obtained on the evaluation environment in the trial phase. Note that the
performance is not guaranteed at the start of its operation.