1 of 20

Software for PED studies

(About) Compute Resource Needs

7th FCC Physics Workshop 2024

Annecy, France

January 31, 2024

G Ganis, CERN-EP

2 of 20

Outline

Quick recap about the problem
How it looks for FCC-ee
A few remarks

G Ganis, S&C, Compute Resources Needs, 7th FCC Physics Workshop 2024, 31 January 2024

3 of 20

Computing and HEP

Cost of IT-related components in HEP experiments, from software development to storage to processing power, constantly raised in the last decades

Starting with 90’s, i.e. when computing has become significant

Experiment workflows and their components have become more and more complex in many directions

Increasing expected data samples, hence increasing needs
Varying scenario of ressources

From single-core to multi-core to heterogeneous processing units
Evolution of storage systems from local to distributed and cache hierarchies

Continuous evolution/optimisation of software and data structures

Estimating/predicting needs for computing resources has become crucial to plan for and secure them
All this requires modeling

G Ganis, S&C, Compute Resources Needs, 7th FCC Physics Workshop 2024, 31 January 2024

4 of 20

Modeling the resource needs

Simple, in principle

Define the needs and the activities to satisfy them

E.g. ‘MC samples for year Y_N’ requires the activities

Event generation, simulation, reconstruction

To each given activity corresponds a set of workflows, each defining a set of resources to be used, e.g.

Event generation on GPU producing output on local NAS
Simulation on the Grid producing output on the Grid
Reconstruction on HLT producing output on EOS

Put everything together to get the compute ressources needs

Taking into account experiment guidelines and policies�

All this depends on assumptions which may be or may be not well defined
For LHC they reasonably well defined

G Ganis, S&C, Compute Resources Needs, 7th FCC Physics Workshop 2024, 31 January 2024

Inspired by D Lange et al, CMS Computing Resources Modeling

5 of 20

Projections of resource needs of HL-LHC

G Ganis, S&C, Compute Resources Needs, 7th FCC Physics Workshop 2024, 31 January 2024

CMS, ATLAS

Better

software

Better software

Pledged +20%

Pledged +10%

Processing power

Storage

6 EB

6 of 20

How does it look for FCC-ee?

G Ganis, S&C, Compute Resources Needs, 7th FCC Physics Workshop 2024, 31 January 2024

7 of 20

Modeling resources for a project “en devenir”

Same reasoning, less precise assumptions, different purpose

Monte Carlo only
No data processing, calibration, …
Several detector concepts
Several digitisation, reconstruction options, …�

Several purposes: e.g.

Projects needs for pledged resources “in production”
Estimate the potential of a limited set of resources to set priorities in the short term

G Ganis, S&C, Compute Resources Needs, 7th FCC Physics Workshop 2024, 31 January 2024

8 of 20

Workflows to support for FCC

MDI codes

MC Generators

Full / Fast Simulation

Digitisation / Reconstruction

Parametrized

simulation

Geometry

MDI format readers

Pileup / MDI overlay

Analysis

MC Generators Interface

Workload and Data Management

Software Infrastructure

(Repositories,Build/Test/Deploy)

G Ganis, S&C, Compute Resources Needs, 7th FCC Physics Workshop 2024, 31 January 2024

9 of 20

What we have now

CERN

EOS volumes

500 TB for central productions (157 TB free, stil used by some CDR files)
200 TB for analysis, starts to be used

CPU: 9000 HS06 on lxbatch
Integrated in iLCDirac

Other sites integrated

BARI
CNAF
Glasgow (storage only)

Some GPU resources

CERN, EuroHPC

G Ganis, S&C, Compute Resources Needs, 7th FCC Physics Workshop 2024, 31 January 2024

Not yet limited, in general, but start reaching the boundaries

10 of 20

A first resources analysis for FCC-ee¹

Assumptions

Nominal luminosities

{90, 12, 5, 0.2, 1.5} ab^-1 at √s = {91.2, 160, 240, 350, 365} GeV

MC reference sample = data sample

i.e. 3x10¹² visible Z decays, 10⁸ WW events, 10⁶ ZH events, 10⁶ tt events

Event sizes (see next)

RAW: 1 - 2 MB/evt
AOB: 5 - 10 kB/evt

Processing Power

CERN Openstack Core = 10-15 HEPSpec06
FCC currently assigned processing units = Computing Unit = 9000 HEPSpec06
CERN OpenStack node used for tests: 16 cores, 32 GB RAM

G Ganis, S&C, Compute Resources Needs, 7th FCC Physics Workshop 2024, 31 January 2024

Based on: GG, C Helsens: EPJ Plus (2022) 137:30

11 of 20

Event Sizes estimations¹

G Ganis, S&C, Compute Resources Needs, 7th FCC Physics Workshop 2024, 31 January 2024

RAW: 1 - 2 MB / evt

AOD: 5 - 10 kB / evt

1. Based on: F. Grancagnolo: Event Rates at Z-pole,� talk presented at 4th FCC PED workshop, Nov 2020

12 of 20

Storage requirements

G Ganis, S&C, Compute Resources Needs, 7th FCC Physics Workshop 2024, 31 January 2024

≈HL-LHC

13 of 20

Computing requirements

G Ganis, S&C, Compute Resources Needs, 7th FCC Physics Workshop 2024, 31 January 2024

2000-3000 years!

14 of 20

Computing estimation remarks

MC event generator can be challenging for full-scale production

Code optimisations and/or filtering techniques might be required

Full simulation times have been cross-checked with ATLAS times for similar multiplicity

Recent Geant4 is up to a factor of 2 faster for ATLAS
Fast simulation techniques (see A Zaborowska talk) might help
Other option: selected simulation, i.e. non simulating what is never touched is not simulated

Reconstruction

Between 10% (ALEPH) - 30% (BELLE) of simulation
Could really benefit from using heterogenous resources

G Ganis, S&C, Compute Resources Needs, 7th FCC Physics Workshop 2024, 31 January 2024

15 of 20

Putting all together …

The FCC-ee resource needs are of the same order of HL-LHC

If HL-LHC solves the problems, FCC-ee gets it for free�

Putting all together

The resources currently available are O(1000) off for full simulation for FSR

Might be ok for parametrized simulation

Numbers should be multiplied by the number of detector variations and analysis

Although some optimisation might be possible

G Ganis, S&C, Compute Resources Needs, 7th FCC Physics Workshop 2024, 31 January 2024

16 of 20

What next: S&C

Investigate possibility to get more resources - e.g. spare cycles from WLCG - and be ready to use efficiently all what becomes available

See L. Valentini, iLCDirac

Increase quality and efficiency of code

Long and expensive process, in general.
New faster simulation techniques promising (see A. Zaborowska talk)

Need to understand how to go beyond the full sim training statistics

Investigate possibility selective/filtered simulations

Filters at generation level
Simulate only parts of relevance

Facilitate interplay targeted full simulation and parametrized simulation

E.g. Automatic/optimal creation of Delphes configurations

G Ganis, S&C, Compute Resources Needs, 7th FCC Physics Workshop 2024, 31 January 2024

17 of 20

What next: Physics Performance, …

Investigate (statistical) technologies to go beyond the rule of thumb� MC Sample = Expected data sample�reducing the number of events required

Could be useful also in perspective, when data will be there

LHC is testing at similar statistics

Use this to identify processes requiring more attention

G Ganis, S&C, Compute Resources Needs, 7th FCC Physics Workshop 2024, 31 January 2024

18 of 20

Final remarks

Available computing resources are limited and do not allow full statistics studies

This is rather normal, given the investment they would represent

Improvements in the code are always possible, but not such to change completely the picture
MC implements our knowledge, should now what to expect
We should try to use all that to identify possible criticalities where to use the available resources

G Ganis, S&C, Compute Resources Needs, 7th FCC Physics Workshop 2024, 31 January 2024

1 of 20

2 of 20

3 of 20

4 of 20

5 of 20

6 of 20

7 of 20

8 of 20

9 of 20

10 of 20

11 of 20

12 of 20

13 of 20

14 of 20

15 of 20

16 of 20

17 of 20

18 of 20

19 of 20

20 of 20