F5C GPU performance

We tested on a variety of different systems (covering embedded systems, laptops, workstations and servers - table below) for comparing the performance of f5c CPU version with the GPU accelerated version.

System name	Info	CPU	CPU cores/threads	RAM	GPU	GPU memory
Jetson TX2	NVIDIA Jetson TX2 embedded module	ARMv8 ARM Cortex-A57 + NVIDIA Denver2	6/6	8 GB	Tegra	shared with RAM
Acer E5 laptop	Acer F5-573G	i7-7500U	2/4	8 GB	Gefore 940M	4 GB
XPS15 laptop	Dell XPS 15 laptop	i7-8750H	6/12	16 GB	Gefore 1050 Ti	4 GB
Old workstation	Unbranded Workstation	i7-6700K	4/8	32 GB	Tesla C2075	6 GB
HP workstation	HP Z640 workstation	Xeon E5-1630	4/8	32 GB	Kepler K40	12 GB
Dell server	Dell server	Xeon Silver 4114	20/40	376 GB	Tesla V100	16 GB

Speedup of adaptive banded event alignment

Adaptive banded event alignment algorithm consumed a vast majority of ~70% runtime of the whole methylation calling processing when run on CPU. Hence, as the first step we decided to offload this alignment process to the GPU. In order to maximise the benefit, we follow a heterogenous approach in our GPU accelerated version - both the CPU and the GPU are simultaneously utilised. The performance of adaptive banded alignment are below.

Overall speedup of processing (excluding I/O)

The overall performance for the whole methylation calling process (not including I/O) is as below.

Overall speedup (including I/O)

The overall performance for the whole methylation calling process (including I/O) is as below.

Note that the speed up for the overall speedup for the Dell server was limited by slow fast5 I/O compared to the processing capability of the Tesla V100.

Note : In all cases the programs were executed with maximum number of hardware threads supported by the CPU of the particular system. The small chr20 dataset (total ~150M bases with ~20K reads) was used.

Acknowledgement : Thanks @GPrathap for running the experiments on the Acer laptop.