How fast is BAYESTAR?

In order to achieve the lowest possible latency for real-time observations, BAYESTAR is optimized to use instruction-level parallelism on modern x86_64 processors and thread-level parallelism via OpenMP to scale up to systems with up to tens or hundreds of cores.

The plot below shows the median run time of BAYESTAR, averaged over all O3 events, for a few selected recent version of ligo.skymap. Code optimizations between 2019 and 2020 resulted in a speedup by almost a factor of 5.

These benchmarks were collected on a Dell PowerEdge C6420 system with two Intel® Xeon® Gold 6136 @ 3.00GHz CPUs and hyperthreading enabled, for 48 hardware threads. This is the same hardware configuration on which BAYESTAR was deployed for online LIGO/Virgo alerts during Advanced LIGO and Virgo’s third observing run (O3).