Download presentation
Presentation is loading. Please wait.
Published bySarah Wilkinson Modified over 6 years ago
1
Reflective Memory vs. Ethernet Evaluating Data Network Hardware Solutions for LCLS Fast Feedback Controls Marya Pearson
2
Introduction The LCLS Fast Feedback project aims to provide a method for controlling the beam parameters and deliver a stable XFEL. Feedback a technique for controlling these parameters collect measurements evaluate algorithms produce instructions for action Physics Requirement Document (PRD) outlines all specifications for a data network that will facilitate feedback.
3
The PRD Describes the parameters we wish to control.
Transverse Longitudinal Launch Here, the time budget is introduced. Data Network requirements Deterministic Scalable Low-Latency Affordable
4
Time Budget According to the Physics Requirement Document (PRD) the machine operates at 120 Hz and thereby the 8.3 ms time budget is derived. 1+1+6 = 8ms 1ms BPMs send data 1ms for fast feedback ~6ms for magnets to react In some cases time budget is adjustable more or less.
5
Time Budget Based on the 120 Hz constraint
The idea is that at least 7ms of the 8.3 ms are fixed. That leaves the feedback 1ms to regulate.
6
Injector Launch IOC:IN20:BP02 BPM9-15 IOC:IN20:MG01
XC04, XC07, YC04, YC07 Loop Rate: 10 Hz This means that the Inj.L feedback can afford 100ms, however to support 120 Hz beam operation the system should operate within 8.3 ms.
7
Data Network Hardware Solutions
Reflective Memory (RM) Star topology Ring topology Ethernet Point-to-Point Multicast
12
Pros Cons Cost (InjL) Latency Scalability Deterministic Other
Reflective Memory-Ring Built-in redundancy Eliminates use of hub No processor involvement in initialization Error detection and bypass Data must pass through each node May lead to difficulty in scaling $13 k 40 ns each node difficult yes Dynamic packet size 4 to 64 bytes Transfer rate Mbytes/s Reflective Memory-Star Processor identifies it as RAM Commercially available Increased scalability Costly option More hardware involved $17 k Error detection Redundant transfer mode Point-to-Point Ethernet Least expensive Direct communication between nodes Not ideal if many devices involved Latency increases exponentially price of additional port Equation: 15n2+85n+3 n= #of nodes Most readily compatible option Multicast Ethernet Selective data delivery Conserves bandwidth One-to-many or Many-to-many Requires significant programming Worst case scenario and recovery $10.5 k +price of additional port 6 µs for 64-byte packets ~4.5 µs for 48 byte packets limited Throughput: 72 mpps
13
Decision and Future Work
Ethernet with Multicast chosen for first prototype simulations must be done to measure performance. Use Cisco products and collaborate with Cisco engineers to set up a test stand.
14
Acknowledgements This research was conducted at the U.S. Department of Energy Stanford Linear Accelerator Center. I would like to express my gratitude to my mentors, Ernest Williams Jr. and Sheng Peng. I am also thankful for the assistance of Farah Rahbar, Susan Schultz, and Stephen Rock.
15
Questions?
Similar presentations
© 2025 SlidePlayer.com. Inc.
All rights reserved.