Post on 23-Oct-2019
PES
CERN IT DepartmentCH-1211 Geneve 23
Switzerlandwww.cern.ch/it
CERNITDepartment
Future of Batch Processing at CERNHEPiX Fall 2013
Jerome Belleman Daniel PekCERN IT
October 2013
PES
CERN IT DepartmentCH-1211 Geneve 23
Switzerlandwww.cern.ch/it
CERNITDepartment
2 – Future of BatchProcessing at CERN
Outline
1 The Future, LSF and its Shortcomings
2 Alternative Batch Systems
3 Initial Results
PES
CERN IT DepartmentCH-1211 Geneve 23
Switzerlandwww.cern.ch/it
CERNITDepartment
3 – The Future, LSFand its Shortcomings
Section 1
The Future, LSF and its Shortcomings
PES
CERN IT DepartmentCH-1211 Geneve 23
Switzerlandwww.cern.ch/it
CERNITDepartment
4 – The Future, LSFand its Shortcomings
Current Setup
IBM LSF 7.0.6
4 000 nodes
SLC5 → SLC6Physical → virtual (≈ 1000 virtual so far)
> 65 000 cores
400 000 jobs/day
±70 000 running jobs
PES
CERN IT DepartmentCH-1211 Geneve 23
Switzerlandwww.cern.ch/it
CERNITDepartment
5 – The Future, LSFand its Shortcomings
A Large, Busy Cluster
#jo
bs
pending —running —
15Ju
ne
29Ju
ne
13Ju
ly
27Ju
ly
10Augu
st
24Augu
st
7Sep
tem
ber
100k
200k
300k
400k
PES
CERN IT DepartmentCH-1211 Geneve 23
Switzerlandwww.cern.ch/it
CERNITDepartment
6 – The Future, LSFand its Shortcomings
Goals and Concerns
Goals Concerns with LSF
30 000 to 50 000 nodes 6 500 nodes max
Cluster dynamism Adding/Removing nodesrequires reconfiguration
10 to 100 Hz dispatch rate Transient dispatchproblems
100 Hz query scaling Slow query/submissionresponse times
PES
CERN IT DepartmentCH-1211 Geneve 23
Switzerlandwww.cern.ch/it
CERNITDepartment
7 – The Future, LSFand its Shortcomings
Addressing Operational Issues
Hired an LSF consultant
Suggested a few minor enhancements:
Kernel parametersMemory limitsCPU bindingLogging and accounting file handling
No miraculous improvements
Taught us a lot
PES
CERN IT DepartmentCH-1211 Geneve 23
Switzerlandwww.cern.ch/it
CERNITDepartment
8 – AlternativeBatch Systems
Section 2
Alternative Batch Systems
PES
CERN IT DepartmentCH-1211 Geneve 23
Switzerlandwww.cern.ch/it
CERNITDepartment
9 – AlternativeBatch Systems
Current Candidates
SLURM 2.5.7
HTCondor 8.1.0
Son of Grid Engine 8.1.3
LSF 8/9
PES
CERN IT DepartmentCH-1211 Geneve 23
Switzerlandwww.cern.ch/it
CERNITDepartment
10 – AlternativeBatch Systems
Batch Testing Framework
JobGenerator
JobSubmitter
Harvester
BatchSystem
Poller DB
[
{ "cmd": "tests.d/psleep.py 120",
"count": 1000,
"instance ": "batch",
"user": "atlas",
"queue": "debug" }
]
PES
CERN IT DepartmentCH-1211 Geneve 23
Switzerlandwww.cern.ch/it
CERNITDepartment
10 – AlternativeBatch Systems
Batch Testing Framework
QueryGenerator
QuerySubmitter
Harvester
BatchSystem
Poller DB
[
{ "cmd": "tests.d/bjobs.py",
"count": 1,
"instance ": "batch" }
]
PES
CERN IT DepartmentCH-1211 Geneve 23
Switzerlandwww.cern.ch/it
CERNITDepartment
11 – Initial Results
Section 3
Initial Results
PES
CERN IT DepartmentCH-1211 Geneve 23
Switzerlandwww.cern.ch/it
CERNITDepartment
12 – Initial Results
Submission Rates
500 1000 1500 2000time in secs
0
50
100
150
200
250
300
350
subm
issi
on r
ate
in H
z
job scaling testSLURM
HTCondor
Son of Grid Engine
PES
CERN IT DepartmentCH-1211 Geneve 23
Switzerlandwww.cern.ch/it
CERNITDepartment
12 – Initial Results
Submission Rates
500 1000 1500 2000time in secs
0
50
100
150
200
250
300
350
subm
issi
on r
ate
in H
z
job scaling testSLURM
HTCondor
Son of Grid Engine
PES
CERN IT DepartmentCH-1211 Geneve 23
Switzerlandwww.cern.ch/it
CERNITDepartment
12 – Initial Results
Submission Rates
500 1000 1500 2000time in secs
0
50
100
150
200
250
300
350
subm
issi
on r
ate
in H
z
job scaling testSLURM
HTCondor
Son of Grid Engine
PES
CERN IT DepartmentCH-1211 Geneve 23
Switzerlandwww.cern.ch/it
CERNITDepartment
12 – Initial Results
Submission Rates
500 1000 1500 2000time in secs
0
50
100
150
200
250
300
350
subm
issi
on r
ate
in H
z
job scaling testSLURM
HTCondor
Son of Grid Engine
PES
CERN IT DepartmentCH-1211 Geneve 23
Switzerlandwww.cern.ch/it
CERNITDepartment
12 – Initial Results
Submission Rates
500 1000 1500 2000time in secs
0
50
100
150
200
250
300
350
subm
issi
on r
ate
in H
z
job scaling testSLURM
HTCondor
Son of Grid Engine
PES
CERN IT DepartmentCH-1211 Geneve 23
Switzerlandwww.cern.ch/it
CERNITDepartment
13 – Initial Results
Scalability
Requirements for measurement:
Piggyback on existing resources
Need to simulate many queries
Experience:
Slow SLURM startup with too many hosts?
HTCondor made to scale out
CPU load on the master(s)
PES
CERN IT DepartmentCH-1211 Geneve 23
Switzerlandwww.cern.ch/it
CERNITDepartment
14 – Initial Results
General Impressions
SLURM HTCondor Son of GridEngine
CurrentLSF
Config(Puppet)
Struggle Comfortable1st time,updatesharder
Needssharedfilesystem
Complex
Maturity Easy tofreeze
Trustworthy Rougharound theedges
Trustworthy
Dynamic Notreally
Yes Yes Not really
Doc Poor Solid Solid Solid
Communitysupport
Ratheruninter-ested
Veryenthusiastic
Enthusiastic Commercial
PES
CERN IT DepartmentCH-1211 Geneve 23
Switzerlandwww.cern.ch/it
CERNITDepartment
15 – Future of BatchProcessing at CERN
Required Features
Grid support
Kerberos/AFS
Accounting
Host normalisation
Fairshare scheduling
Support for commercial applications
IPv6?
PES
CERN IT DepartmentCH-1211 Geneve 23
Switzerlandwww.cern.ch/it
CERNITDepartment
16 – Future of BatchProcessing at CERN
Conclusions
Replacement candidates:
SLURM feels too young
HTCondor mature and promising
Son of Grid Engine fast, a bit rough
What’s next:
Host scalability
Query load
Features
PES
CERN IT DepartmentCH-1211 Geneve 23
Switzerlandwww.cern.ch/it
CERNITDepartment
17 – Future of BatchProcessing at CERN
Thanks!
Questions?