Research topic and goals
Many-core architecture is widely used in clusters to achieve power-efficient high performance systems. As the number of computing cores in a system increases, traditional data parallelism with threads cannot achieve high performance because of load-imbalance between cores and large synchronization overhead. Researchers are investigating task parallelism as a promising solution for both load balancing and synchronization overhead. Several programming models are being developed to integrate task programming models into node-level programming models to establish a hybrid programming model for many-core clusters.
RIKEN AICS and University of Tsukuba have been developing their PGAS language, XcalableMP, and working on a new feature that enables thread-level tasks on distributed memory systems. Task programming in XcalableMP is independent from other programming models such as OpenMP, and data transfer should be specified explicitly using the PGAS feature so that the user can expect the performance and optimize it easily. OmpSs have been being developed by Barcelona Supercomputing Center (BSC). It extends the standard OpenMP task model to support advanced asynchronous parallelism and heterogeneous programming for accelerators such as GPUs. OmpSs provides interoperability with MPI for hybrid parallelism and its cluster version enables multi-tasking on clusters based on a shared memory space.
Although two programming models take different approaches, they can share underlying software layers such as communication runtime and thread generation and scheduling mechanics. We propose a collaboration between RIKEN AICS and BSC to share OmpSs2 runtime as a common infrastructure for multi-tasking on large-scale many-core clusters. Since RIKEN AICS have been mainly focused on its PGAS runtime and BSC has been developed thread programming infrastructure, we are interested in extending OmpSs2 runtime to support PGAS operations. Exchanging experiences and technology can boost the development on both sides, and enable open programming interface which can lead future collaboration.
Results for 2017/2018
This activity has just started in March 2018. No results yet.
Results for 2018/2019
We have implemented a Blocked Cholesky factorization code with three schemes. The codes emulate our XcalableMP task runtime. The first one (ver1) is to implement it with multiple thread that communicated with each other. The second scheme (ver2) is to introduce a dedicated communication thread. For comparison, we implemented the code with OmpSs (ver3). A Blocked Cholesky factorization code is used for the performance evaluation. With 32 nodes of a Intel Knights Landing cluster (Oakforest-PACS in the University of Tokyo), ver2 achieves 8 TFLOPS, which is better than ver1 and 3 (7 TFLOPS both). On a Intel Xeon platform (COMA in University of Tsukuba), ver3 shows lower performance than ver1 and ver3 because of the wait overhead in OmpSs runtime.
Visits and meetings
To be planned.
Impact and publications
The first steps will be using OmpSS runtime as a XMP runtime. We will investigate how to combine PGAS operations onto OmpSS runtime.