Speaker
Description
We present a case for ARM chips as an alternative to standard x86 CPUs at WLCG sites, as we observed better performance and lower power consumption in a number of benchmarks based on actual HEP workloads, from ATLAS simulation and reconstruction tasks to the most recent HEP-Score containerised jobs.
To support our case, we present novel measurements on the performance and energy consumption of two comparable machines (ARM and x86), including the efficiency of single- and multi-thread jobs, and the effect of hyper-threading and clock throttling as a means to reduce energy demand at specific times of the day.
We will complement these results by summarising our experience in enabling ARM resources at our Tier2 site, and by showing the current state of ARM support from LHC experiments.
This will help other Grid sites justify the procurement of ARM machines, as opposed to x86, when they have a choice.
Moreover, we have explored alternative solutions for making energy savings at typical WLCG sites, including optimization of the way HTCondor allocates jobs to worker nodes, as well as load shaping as a means to reduce power consumption during peak hours.
Consider for long presentation | Yes |
---|