Life Sciences
Information and Communication
Nanotechnology / Materials
Manufacturing Technology
Human and Social Sciences
Energy
Environment
Tourism / Community development
Arctic Research
Social Infrastructure
Open Facilities
- Home
- Information and Communication
- Communication-avoidant Matrix Calculation Algorithm
Communication-avoidant Matrix Calculation Algorithm
Research and development of algorithms suitable for massively parallel computers
With the spread of massively parallel computers, it has become increasingly important to reduce the communication time associated with parallel computation. In this research, we aim to improve the performance of matrix computation algorithms by using an approach called “Communication Avoiding (CAA).”
Content of research
With parallel processing using large-scale parallel computers, the data communication time is often more important than the computation time. In particular, the large communication latency (the cost incurred regardless of the amount of data to be communicated) has become a problem, and there is a strong need to reduce the communication frequency (communication avoidance). We are reviewing existing matrix computation algorithms from the viewpoint of communication avoidance, and are researching and developing new algorithms for massively parallel computers that reduce the communication frequency.
-
Our communication-avoiding matrix computation (QR decomposition) algorithm:
Performance evaluation results using the K computer at RIKEN have confirmed the effectiveness of the system.
Potential for social implementation
- ・All simulation or data analysis using parallel computers that requires matrix calculations (e.g., eigenvalue and singular value calculations).
Appealing points to industry and local governments
We can give advice and introduce existing libraries concerning problems not only on this research subject and parallel computers, but also on matrix computation in general.
Inquiry about this research
