Reports and Measurements

This page serves to collect reports and measurements on the GMT.

Booster Test

• Timing events during 1st Booster-Mode Test (2021-12-17)
• 2nd Booster-Mode Test (2021-12-22)

Timing Receivers

• saftlib
  • Saftlib latency; including fallout vs enigma (2021-02-18)
  • Saftlib 2.0 latency improvement (2019-02-13)
  • Saftlib/Etherbone/LM32 latency measurement (2017-10-03)
  • ECA and Saftlib: Torture report about GMT with Debian on PC and SL6/CentOS 7 on SCU (2016-07-22)
• etherbone
  • Etherbone Performance Measurements (2018-01-31)
• FPGA
  • Pseudo SRAM Acess from lm32 (2019-02-21)

White Rabbit Network

• Performance evaluation of the GSI timing network using XenaBay (2024-01-24)
• WR PTP synchronization of WRS under excessive PTP and LLDP traffic (2022-05-27)
• White Rabbit Switch VLAN Tests (2020-09-29)
• Latency and Loss of Timing Messages in the Timing System (2019-12-30)
• WR-ZEN (2019-09-11)
• Latency Test Report of the WR Network (2016-05-12)
• RFC 2889 Network Test (2016-08-25)
• Irradiation of Radiation Hard Fibres at SIS18 (2015-07-24)
• Irradiation of Radiation Hard Fibres at HHD (2014-01-23)

MPS Tests

• Network and node performance measurements (2022-03-28)

Other

• SCU Kernel Task Switching Latency (2018-09-03)

The Unofficial Feature List (Out of Date)

Although the subject is extremely complex, it is tried to summarize the status of some key characteristics of the GMT in the following table. It is intended to give safe (worst case) values of the actual performance including a reference; same for the specification. Values are taken at various conditions that might not be representative.

Sub-System	Item	Stat	Reported Value	Specified Value	Date Rep (Spec)	Ref Rep (Spec)	Remark
OS
	CentOS RT: task switching latency		100 us (ub), 20 us (av)	< 1 ms (?)	09/2018 (<2010 ?)	priv comm. from HEL (SR)	measurement on SCU using preempt-test, only 3us gain without hyperthr
Timing Receiver
	FTRN userland latency @ API		93 us (ub), 62 us (av)	< 1 ms (?)	12/2012 (<2010 ?)	article (N/A)	1 ms spec is overal FEC performance
	FTRN userland latency @ saftlib		2500 us (ub), 1400 us (av)	< 1 ms (?)	11/2017 (<2010 ?)	wiki (N/A)	1 ms spec is overal FEC performance, ok for 2018
	FTRN output (Lemo) jitter (RMS)		20 ps (av)	100 ps	11/2014 (01/2013)	article (F-DS-C-06e)	'term jitter' defined in F-DS-C-05e
	FTRN output (Lemo) precision (RMS)		50 ps (av)	1 ns	11/2014 (01/2013)	article (F-DS-C-06e)	'term precision' defined in F-DS-C-05e
	FTRN output (Lemo) accuracy (RMS)		ok, needs re-checking with every release	2-3 ns	11/2014 (01/2013)	wiki (F-DS-C-06e)	'term accuracy' defined in F-DS-C-05e
Timing Network
	upper bound latency		< 390 us @ 4 WRS layers	500 us [1]	06/2016 (2012 ?)	report (minutes, wiki)	F-DS-C-05e mentions '200 us' as 'should'
	throughput (messages)		> 16 messages @ 500 us (sl)	16 msg. @ 500 us (sl)	12/2017 (12/2017)	Dry Run #3 ( minutes )
	throughput (frame size)		< 60 MBit/s	100 MBits/s	08/2016 (09/2014)	report (wiki)	1 to N ports @ 64 bytes/frame, ok for 2018
	loss rate		< 2e-8 @ 600 Hz msg. rate	1e-12	08/2016 (01/2013)	wiki (F-DS-C-05e)	without forward error correction, ok for 2018
	loss rate		> 90% @ > 10% bandwidth	1e-12	08/2016 (01/2013)	wiki (F-DS-C-05e)	for small packets of 64 and 128 bytes. Better for larger packets.
	max numbers of WRS layers		4	5	06/2016 (2012 ?)	report (various minutes)	maybe 4 layers are sufficient
Data Master @2018
	upper bound latency per lm32 CPU		< 500 us @ 16 msg (sl)	500 us [2] (sl)	12/2017 (12/2017)	minutes (./.)	update 2018-sep-03: max reported upper latency is 505 us
	throughput per lm32 CPU		> 16 messages @ 500 us	16 msg. @ 500 us	12/2017 (12/2017)	minutes (./.)
	capacity per lm32 CPU		600 msg @ 2018	600 msg @ 2018	12/2017 (12/2017)	minutes (./.)	1 control block 'eats' 11msg
	number of patterns lm32 CPU		4 @ 2018	> 10 ?	12/2017 (12/2017)	minutes (N/A)	not really specified; 4 patterns ok for 2018
	number of lm32 CPU		3	8 ?	12/2017 (2012 ?)	minutes (N/A)	not really specified; 3 cores ok for 2018

Table: Status of key characteristics of the GMT. Some values are marked by upper bound (ub), or average (av) or sliding window (sl).

[1] The Data Master sends messages to the White Rabbit network 500 us ahead of their scheduled time. The budget of 500 us includes the processing time in the timing receiver. [2] The upper bound latencies of the network and the data master must be added. It was decided to use a total upper bound latency of 1 ms for the 2018 beam time. Update: The '1ms upper bound' is kept for the 2019 und 2020 beam times.

-- DietrichBeck 22 April 2024