1
EIC Computing Meeting
Oct. 18 2021
D. Lawrence
Computing Plan Status
Data Storage and Compute
2
ECCE Bi-Weekly meeting - Computing Team Report - Sep. 27, 2021
Detector
FEB (digitization)
Online Buffer
EBDC (few days)
Online Event Filter
CPU, FPGA, GPU
Offline Buffer
(few weeks)
Calibration
raw
storage
Reconstruction
recon
storage
Experimental Hall and Counting House (Project Funds)
Data Center(s): SDCC [,JLab, …]
(Operations Funds)
HTC Compute Facilities
SDCC ,JLab, …
(Operations)
FELIX-like
FEP (initial filter)
O(100Tbps)
O(10Tbps)
O(0.1Tbps)
O(0.1Tbps)
O(10GB/s)
O(1GB/s)
O(0.1Tbps)
O(0.01Tbps)
reduce ~x2
mitigate risk of network failure
Data Storage and Compute
3
ECCE Bi-Weekly meeting - Computing Team Report - Sep. 27, 2021
Detector
FEB (digitization)
Online Buffer
EBDC (few days)
Online Event Filter
CPU, FPGA, GPU
Offline Buffer
(few weeks)
Calibration
raw
storage
Reconstruction
recon
storage
Experimental Hall and Counting House (Project Funds)
Data Center(s): SDCC [,JLab, …]
(Operations Funds)
HTC Compute Facilities
SDCC ,JLab, …
(Operations)
FELIX-like
FEP (initial filter)
O(100Tbps)
O(10Tbps)
O(1Tbps)
O(0.1Tbps)
O(10GB/s)
O(1GB/s)
O(0.1Tbps)
O(0.01Tbps)
Offline Event Filter?
O(1Tbps)
reduce ~x2
mitigate risk of network failure
EIC Streaming Readout (From Fernando Barbosa’s talk at AI4EIC Sep. 9, 2021)
factor of 100 in data reduction
from EIC Yellow Report
From YR 2.10
“Most of the key physics topics discussed in the EIC White Paper [2] are
achievable with an integrated luminosity of 10 fb−1 corresponding to 30 weeks of
operations.”
10x1033 cm-2s-1 x 30 weeks
=
18.1 fb-1
(assumes 100% beam up time. YR also says 60% operational efficiency which would bring this to 10.3 fb-1)
bottom line: 10fb-1 represents roughly first year of running with about half of that for ep scattering
For ep scattering representing half of a 30 week year of running we expect:
50μb x 5fb-1 = 250B collisions
Inclusive ep scattering MC DST: ~600MB/2k events = 300kB/event
250B/(15weeks*7days/wk*24hr/d*3600s/hr)/(60% accel. efficiency) =46kHz
46kHz * 300kB = 14GB/s (140Gbps)
from EIC Yellow Report
close to Jin’s simulation calculation of O(100Gbps)
Multiple levels of event filter
| Input/Output | Reduction Factor | Technology |
FELIX-like Compute Interface | 100Tbps/10Tbps | 10 | FPGA |
Online Event Filter | 10Tbps/1Tbps | 10 | FPGA, (GPU), CPU |
Offline Event Filter | <1Tbps/01.Tbps | ~10 | FPGA, GPU, CPU |
Reconstruction | 1Tbps/0.1Tbps | 10 | (FPGA), GPU,CPU |
Federated Computing Model
Offline Buffer
Calibration
raw
storage
Reconstruction
recon
storage
Site 1: (e.g. SDCC)
Offline Buffer
Calibration
raw
storage
Reconstruction
recon
storage
Site 2: (e.g. JLab)
Offline Buffer
Calibration
raw
storage
Reconstruction
recon
storage
Site 3: (e.g. NERSC)
Detector
FEB (digitization)
O(100Tbps)
O(10Tbps)
O(1Tbps)
FELIX-like
FEP (initial filter)
Online Event Filter
CPU, FPGA, GPU
Online Buffer
EBDC (few days)
Offline Event Filter
O(0.1Tbps)
reduce ~x2
mitigate risk of network failure
Federated Computing Model
Benefits of using a federated model in the near-time pipeline�
Worldwide LHC Computing Grid (WLCG)
An EIC Butterfly Model
EIC
JLAB
SDCC
BATES
NERSC
OSG
University
University
University
University
University
University
University
University
Tier 0
Tier 1
Tier 2
Tier 3
Tier 3
Nearly all storage (raw data, reconstructed data, simulated data) is stored across Tier 1 sites
ECCE Computing Plan
12
ECCE Bi-Weekly meeting - Computing Team Report - Sep. 13, 2021
Key features:
Open Questions
Backups