Automating Data Layout Conversion in a Large Cosmological Simulation Code

•Gkernel is: – isolated representative Gadget code kernel (“halo finder”) – stand-alone application, avoids simulation overhead • Node-level optimization study [3] in the frame of the IPCC [4] • Target computing systems: – Knights Corner (KNC), Ivy Bridge (IVB) – Haswell (HSW), Broadwell (BDW), Knights Landing (KNL) •Main changes: – Data layout optimization – Better threading parallelism Figure 2: Tests on one-socket Xeon systems; 240 threads (4 thr./core) for KNC; 128 threads (2 thr./core) for KNL. Performance improvement: up to 19x faster on KNL; 13.6x on KNC, ca. for 2-5x on Xeon.