1. “L1 farm: some naïve consideration” Gianluca Lamanna (CERN) & Riccardo Fantechi (CERN/Pisa)

2. The L1 farm can’t be avoided The L1 farm can’t be avoided: 100 kHz – the NZS LKr must be read at 100 kHz rare decay junk L2 farm – We are looking for a rare decay: most of the events passing the L0 are junk, avoid to build a huge and complex L2 farm to fully reconstruct events that have to be reject GTK readout window – Some detector (i.e. GTK) needs additional information to reduce the readout window (and then the bandwidth) L1-L2 PCs mixing It’s better to avoid the L1-L2 PCs mixing: protocolrate – Characteristics of the data links (protocol, rate, etc.) – Specialization of the L1 PCs with respect to the sub- detector

3. How to calculate the number of PC in the L1 farm? “type”: Two different “type”: participatingL1 trigger – Detector participating to the L1 trigger decision  number of PC driven by computing power needed to calculate the primitives not participating L1 trigger – Detector not participating to the L1 trigger decision  number of PC driven by data rate and bandwidth

5. Not participating to L1 Two relevant parameter: – Bandwidth10 Gb – Bandwidth: with 10 Gb links only the GTK needs 2 ports (2 PCs?) – Maximum data rate1 MHz ?more events in one packet – Maximum data rate: 1 MHz packet rate isn’t sustainable from an ethernet card (max 200 kHz?)  more events in one packet 1500 B300 B 51MHz/5=200 kHz The number of event per packet depends on the size of the event: for instance assuming standard ethernet packet (1500 B) and RICH event size of 300 B, a factor of 5 can be applied to the max rate (1MHz/5=200 kHz) Jumbo frame 9000 B – Gain by using Jumbo frame (9000 B): buy switches capable to manage Jumbo frames

6. Not participating to L1 1 Gb/s 1500 B 12 us 80 kHz But the maximum frequency of packets in the cable (1 Gb/s) depends on the length of the packet: 1500 B  12 us  80 kHz – Increase the number of PC to cope with high rate 5 events 80 kHz10 ports ? ? – RICH: 5 events per packet  80 kHz  10 ports (2-3 PCs ? ), multicore (100 kHz of interrupts per core?) To evaluate the correct number of PC for each subdetector we need to know: – Events size PPSNIC – Maximum PPS of the NIC interrupts rate – Maximum interrupts rate per core – Number of events per packet

7. Participating to L1 L1TP Some detector should produce primitives for L1TP NM In this case the number of PC should be increased to allow the online data processing: 1 MHz  1 us  N (number of cores) = M (actual time of the algorithm) RICH 100 us 100 cores 10-12 PC For instance: if the RICH algorithm to reconstruct ring will take 100 us  100 cores (10-12 PC) GPUs 2-31 PC 2 video card Using GPUs this number could be highly reduced: RICH 2-3 us  1 PC with 2 video card (probably less)

8. Final message L1 farm The dimension of the L1 farm should be evaluated with care actual rate computing increasedecrease sub-L1farm sub-L1farm L0->L1 farm 10 Gb to 1 Gb 10 Gb to 10 Gb The real dimension depends on actual rate and computing needed : it’s very important to have the possibility to increase and decrease the number of PCs in each sub-L1farm  include small switches for each sub-L1farm (or assure that the L0->L1 farm are large enough for future upgrade)  it’s convenient to use 10 Gb to 1 Gb switches instead of 10 Gb to 10 Gb ???

9. Time measurement one pc to another We are trying to measure transit time sending data from one pc to another using the oscilloscope. LPT Signals on LPT Jitter few us Jitter of the method measured (under control… few us) interrupts frequency data rate Study of time performance as a function of the interrupts frequency and data rate. “gun” TELL1/TEL62 Future step: replace the “gun” with a TELL1/TEL62 (with a signal produced by the firmware on the test connector) PC1: gun PC2: target Interface to LPT