Understanding and Coping with Hardware and Software Failures in a Very Large Trigger Farm
Jim Kowalkowski
Abstract
When thousands of processors are involved in performing event filtering on a trigger farm, there is likely to be a large number of failures within the software and hardware systems. BTeV, a proton/antiproton collider experiment at Fermi National Accelerator Laboratory, has designed a trigger, which includes several thousand processors. If fault conditions are not given proper treatment, it is conceivable that this trigger system will experience failures at a high enough rate to have a negative impact on its effectiveness. The RTES (Real Time Embedded Systems) collaboration is a group of physicists, engineers, and computer scientists working to address the problem of reliability in large-scale clusters with real-time constraints such as this. Resulting infrastructure must be highly scalable, verifiable, extensible by users, and dynamically changeable.
Create a lesson
Related papers
PixelFlow: Token-Level Workload Management for Efficient Distributed DiT Serving
Zhexiang Zhang, Minchen Yu, Yifan Sun et al.
A Kubernetes-Native Request Router for Quality-Aware Inference Serving in the Computing Continuum
Ignjat Karanovic, Pantelis A. Frangoudis, Ivan Čilić et al.
Accelerating Sharded Data Parallelism at Scale with Federated Learning
Gianluca Mittone, Marco Aldinucci
Distributed Edge Inference: an Experimental Study on Multiview Detection
Gianluca Mittone, Giulio Malenza, Marco Aldinucci et al.
P-GADMM: Parallel Group-Based ADMM for Asynchronous Optimization in Heterogeneous Edge Networks
Gaiguo Wei, Qingying Zhang, Heqiang Wang et al.
VERA: Reinforcement Learning for Dynamic Memory Scaling of HPC Workloads in Kubernetes
Ade Pramono, Jie Ren, Ivy Peng