View
2
Download
0
Category
Preview:
Citation preview
CMS StatusTommaso Boccali (INFN Pisa / CERN)Markus Klute (MIT)
1
Outline
● General status - since last LHCC● Resource request for 2020● News from the project and AoB
2
Since last LHCC: HI endgame
● We were in HI data taking○ Concluded on Dec 3rd○ Final statistics is 4.4 B full format
Minimum Bias events + 250 M “physics triggers”
○ Final data size is 5.3 PB RAW
● Prompt processing went on @ T0 (with HLT when available - slide 5), and was done on Jan 7th
● 2nd full copy of RAW @ FNAL by Dec 28th (1 month earlier than expected due to > 2x speed of transfer);
● AODs on Disk (Vanderbilt T2) and Tape (IN2P3 T1) in sync with the production process 3
Since last LHCC: HI endgame
● All in all pretty much as expected (and tested w/IT!)● In the last days Data Taking pushed at 7.3 GB/s
○ 6 GB/s was expected to be the safety limit
● Processed data is in the hands of the analyzers since early January; better than plans which were pointing to end of Jan
4
Processing from last LHCC - HLT utilization
● We had the HLT partially available for the whole period○ Partially: some complete shutdowns due to P5 work, mostly at reduced capacity due to limited
cooling
● Bottom line: whenever technically available (“on”), we can use HLT
HLT processing HI @ T0 HLT processing offline workflowsNo availability (P5 activities)
HLT needed by DAQ (beam on)
5
MC campaigns
● MC 18 (targeted to Moriond19 analyses):
○ 8.2 B done; M19 high priorities fully processed and available
○ No showstopper for analyzers!
● PU200 production for MTD TDR:
○ Admittedly, we started late (end of Jan); >1 month of delay for sw
○ 10M Events the reduced request○ Processed @ 80% level in 1 week
Restart after the ECAL issue
8.2 B events ready for analysis
Feb 2nd
Feb 8th
6
Parking B
● As already reported in Nov, 12 B events collected, ~ 10 B of B hemispheres via a specialized trigger reaching almost 6 kHz
● Since then:○ Specific electron low Pt reconstruction
ready○ Started staging RAW from tape○ Plan is to process ALL the events + skim
O(100M) events of particular interest
● The low Pt electron reconstruction will enter default workflows in the ultra legacy reprocessing
/ParkingBPH*/Run2018*-v1/RAWTOTALS:Datasets: 23Size on Tape: 7596 TB (single copy)Events: 11878715564 Avg Size 639528 bytes / event
7
The Ultra Legacy reprocessing
● Preparation currently as planned without major delays
● We will probably use the sequence○ 2017 → 2018 → 2016
● <PU> is in line with modelling (35)● HLT should be fully available by April● IT is positive that in the same time
scale the P5 ←→ IT link should be upgraded from 4x40 to 4x100 Gbits
○ Over Xmas we touched for a few hours the limit
○ We expect at least 30% additional HLT available during UL
8
NanoAOD evolution & NERSC
● Commissioning is going on, and a number of Moriond analyses will use NanoAOD
● Eventually, we processed all 2016-2018 DT and MC to have a consistent set for RunII, even before Ultra Legacy
● Processed NANOAOD* in 2018:● MC: 45 B events● DT: 35 B events
… using < 10% of the computing resources …
we can scale much higher
● NERSC is used in production at high levels, until grant limits
● Last 90 days shown below● Xeon (20%) and KNL (80%) seamlessly
in production at the same moment● Slower cores (4x) compensated by
larger MT (4x), possible due to the excellent MT capabilities of CMSSW FW
50kCores
20kCores
NERSC # cores
9
Storage status
● The ParkingB and Minimum Bias HI streams were unexpected / reduced in the 2018 modelling when resources were asked
○ CMS was able to handle these without any resource increase (current/future), via special running modes and aggressive cleanings
● As of now:○ T0 tape nearly full, but no additional expected pressure
during LS2○ T1 Tape ok○ Disk areas in
good shape (70% “unmovable”)
T0 tape
T1 tape
10
2019 Production Plan
Nov Dec Jan2019
Feb Mar Apr May Jun Jul Aug Sep Oct Nov Dec Jan2020
Feb Mar
Prompt 10_3
10_5_0 10_6_0
MC for PbPb 10_3
HI-park RECO 10_3 B-Parking RECO
Jamboree 17 JEC SF 17
Jamboree 18 JEC, SF 18
Rereco 2017MC for 2017 physics 10_6
MC UL prep 10_5, for the 3 years, for PF Calibration.
MC Alca prep 17
data AlCa prep 2017
10_5 valid
valid
valid
Rereco 2018MC for 2018 physics 10_6
MC Alca prep 18
data AlCa prep 2018 valid
valid
Rereco 2016MC for 2016 physics 10_6
MC Alca prep 16
data AlCa prep 2016 valid
valid
Jamboree 16 JEC, SF’ 16
MTD TDR
2016 re-MiniAOD/NanoAOD
Still not started (waiting for final sw)
All needed pieced of software and condition are integrated together. Start RelVals production for validation
11
2020 Requests
● CMS confirms a request of +0% (wrt 19) in all the areas
● CMS planned activities in 2020:○ Tails of the UL○ Analyses as in 2018○ RunIII preparation, including a dress rehearsal of
the new sw components■ Rucio, DD4HEP, CRIC, …
● For 2021, CMS updated its modelling in view of recent LHC planning; with respect to 2019 the up-to-date figures are
○ CPU: +30%○ Disk: +20%○ Tape: +25%○ (all to be confirmed in Fall’19) - used to be a flat
+30% 12
News from software and AoB
● CMSSW_10_5 contains the first commits of the new CMS Heterogeneous framework
○ To be finalized in 10_6_0○ See next slide
● The development and test of Rucio, DD4Hep and CRIC continue as planned○ CRIC has substituted SiteDB in production○ Rucio generally on schedule with the plan sketched on Jul 18
■ Next one is “Million file test with Rucio & PB scale tests” - March 19; under preparation
● CB122 approved the move to Apache2 license
13
Goals and status for heterogeneous CMSSW● Identify a general pattern for all accelerators
○ Keep CPU busy while running algorithms on the accelerators○ Achieved with ExternalWork extension to EDProducers○ Framework reuses the thread for other modules while waiting
the external work to finish (== no locks)● Use a single CMSSW configuration for all hardware combinations
○ Achieved with SwitchProducer configuration entity○ Discover available hardware at cmsRun start time, and use
the appropriate EDProducers (can be overridden by config, etc…)
○ Files created on different hardware can be used homogeneously
● Model for NVidia GPUs with CUDA is being prototyped as part of the Patatrack project
○ Supports chains of EDProducers passing data on GPU memory
○ Data transferred to CPU only when needed○ Expose as much of the parallelism as possible
● Same or similar components are expected to be useful for other GPUs, FPGA etc.
○ E.g. SONIC project is using the system for connecting to remote FPGA resources for ML inference
14
Conclusions
● CMS has been able to digest smoothly the large HI dataset taken at the end of 2018; the reconstruction of the ParkingB dataset is imminent
● Production activities have already delivered a complete set of processed DT and MC for the full RunII, in due time for Moriond analysis preparation
● Plans for 2019, including the UL, are progressing on schedule
● CMS confirms the “+0%” requests for 2020, and has slightly reduced the 2021 modelling wrt to previous iterations 15
Recommended