White Papers & Research Articles

System Hardware

Acceleration of Sparce Matrix Multiplication For HPC Application by Sanjiv K Mangal, Sunil Puranik, Milind Athavale and Shreenivas Sapre 

Abstract: This paper presents a scalable design of FPGA based accelerator used to speed up double precision, sparse matrix vector multiplication (SpMV). SpMV operations are difficult to accelerate on standard CPUs due to limited memory bandwidth, poor cache re-use and low compute-communication ratio. SpMV is a good candidate for acceleration by FPGAs as they provide fine-grained parallelism combined with a large number of user configurable I/O pins. We address the problem of SpMV multiplication for large matrices using commodity DRAMs. The design has been targeted for the high-end Xilinx LX330T FPGA and aims to accelerate performance for large matrices generally encountered in practice. Our accelerator delivers performance of 14 GFlops for unstructured matrices, comparing well with related art. 

Computational Fluid Dyanamics

Residence Time Distribution in Steel Tundish using OpenFOAM by Santhosh Kumar and Kishor Nikam  

Abstract: The flow distribution and residence time distribution (RTD) is studied numerically in steel tundish with a representative water model. The Numerical simulations are carried out by an open source CFD software, OpenFOAM. The RTD results of OpenFOAM are compared with experimental and commercial CFD software i.e, Fluent. The dynamics of flow in representative water model of a steel tundish are well captured by numerical simulations. The velocity distribution obtained by OpenFOAM is in very close agreement with Fluent and closely captures the experimentally observed short circuiting phenomena

System Software

External Factors Affecting Performace of MPI by Kiran Nalawade and Shrirang Karandikar

Abstract: A large number of factors influence the performance of an MPI application, such as choice of the MPI library, distribution of ranks to nodes, etc. In addition, the scalability of each application varies across domains and implementation. These factors can cause a variation of more than 100x