Task 1.2. Design of Scalable SoCs for accelerating Gibson and Bruck’s Next Reaction Method
In this task we designed a new parallel architecture for Gibson and Bruck’s NRM-SSA. Although the NRM SSA has a clear performance advantage over Gillespie’s FRM SSA algorithm it is a much harder algorithm to parallelize. As a consequence, it is not surprising that prior to this work there existed only a couple of FPGA implementations of the NRM, which did not scale well beyond N=4 PEs.
We have managed to:
- Design an efficient fully pipelined NRM-SSA PE architecture whose Reaction Cycle latency depends not on the number of reactions in the biomodel but on the average number of dependent reactions, i.e. on the average outdegree of the Dependencies Graph (DG) of the model and not on the number of DG nodes.
- Avoid altogether in our NRM-SSA PE architecture the hardware implementation of a priority queue, that would render it inefficient, and prevent scalability issues faced by previous solutions.
- Reuse as much as possible the hardware components developed for the FRM-SSA SoCs (after appropriate modifications) to minimize PE redesign time.
- The NRM-SSA PE can be embedded into an NRM-SSA core which can be replicated for high performance giving rise to a multiprocessing NRM SoC with excellent scalability characteristics.
- Develop an IP core (in VHDL) for the NRM-SSA SoC that is fully parametric and supports all the capabilities, discussed in Task 1.1, for the FRM-SSA SoC.
Deliverables (technical report)
D1.2 Parameterized HDL SoC descriptions for the NRM-SSA
Evangelos Koutsouradis, George Provelengios, Elias Kouskoumvekakis, Elias S. Manolakos:
Scalable FPGA Accelerator of the NRM Algorithm for Efficient Stochastic Simulation of Large-Scale Biochemical Reaction Networks. In Proceedings of 2015 Euromicro Conference on Digital System Design, Madeira, Portugal, August 26-28, 2015. IEEE Computer Society 2015, ISBN 978-1-4673-8035-5, DSD2015: 583-590