Transcript Wide-Area Xrootd testing at MWT2

Wide-Area Xrootd testing at MWT2
Sarah Williams
Indiana University
1
US ATLAS Tier2/Tier3 Workshop at
UChicago 2009-08-19
Background & Motivation
Currently MWT2_IU and UC function as separate
clusters, using dCache on the SE
Cluster Unification – more storage for analysis,
more capacity available, one set of
configurations to manage
Looking for something easier to manage
Good read, write and analysis performance
Doesn’t introduce new stability or reliability
issues





2
US ATLAS Tier2/Tier3 Workshop at
UChicago 2009-08-19
The Current Situation
3
US ATLAS Tier2/Tier3 Workshop at
UChicago 2009-08-19
Obstacles to a Cross-site Solution
4
US ATLAS Tier2/Tier3 Workshop at
UChicago 2009-08-19
Test Cases
Can a private-ip worker node read & write to a
public-only dataserver? a public-only
redirector?
Does a dual-homed public/private ip dataserver
handle requests from public and private ips
correctly? Does a public/private ip redirector?
What is read & write speed compared to dCache
with dccp?





5
When connecting to a high-performance storage
node? To a combination worker/storage node?
When connecting between sites with xrdcp vs within
the site with dccp
US ATLAS Tier2/Tier3 Workshop at
UChicago 2009-08-19
Test Setup 1
Role
Hostname
Location
Network
Speed
Redirector
Iut2-s1
IU
Public
10G
DataServer
Iut2-s3
IU
Public
10G
Iut2-c034
IU
Public
1G
Uct2-s3
UC
Public/Privat
e
10G
Uct2-c034
UC
Public/Privat
e
1G
Iut2-c042
IU
Public
1G
Uct2-c035
UC
Private
1G
Worker node
6
US ATLAS Tier2/Tier3 Workshop at
UChicago 2009-08-19
Test 1 Conclusions
Can a private-ip worker node read & write to a
public-only dataserver? a public-only
redirector? YES and YES. Switch ACLs must
allow traffic from all nodes to port 1094
Does a public & private ip dataserver handle
requests from public and private ips correctly?
YES


7
US ATLAS Tier2/Tier3 Workshop at
UChicago 2009-08-19
Test Setup 2
Role
Hostname
Location
Network
Speed
Redirector
Uct2-s1
UC
Public/Privat
e
10G
DataServer
Iut2-s3
IU
Public
10G
Iut2-c034
IU
Public
1G
Uct2-s3
UC
Public/Privat
e
10G
Uct2-c034
UC
Public/Privat
e
1G
Iut2-c042
IU
Public
1G
Uct2-c035
UC
Private
1G
Worker node
8
US ATLAS Tier2/Tier3 Workshop at
UChicago 2009-08-19
Test 2 Conclusions
Does a public & private redirector handle
requests from public and private ips correctly?
YES
Switch ACLS must allow traffic from the
redirector to all high-numbered ports on the
data servers


9
US ATLAS Tier2/Tier3 Workshop at
UChicago 2009-08-19
Read Measurements
120
100
80
Max (MB/s)
60
Min
Average
40
20
0
dccp 100MB
10
dccp 1GB
xrdcp 100MB
US ATLAS Tier2/Tier3 Workshop at
UChicago 2009-08-19
xrdcp 1GB
Write Measurements
70
60
50
40
Max (Mb/s)
Min
30
Average
20
10
0
xrdcp 100MB
11
xrdcp 1GB
dccp 100MB
US ATLAS Tier2/Tier3 Workshop at
UChicago 2009-08-19
dccp 1GB
Read Performance by Hardware
120
100
80
Max (MB/s)
60
Min
Average
40
20
0
12
Worker
HP
Worker
HP
100MB
100MB
1GB
1GB
US ATLAS Tier2/Tier3 Workshop at
UChicago 2009-08-19
Write Performance by Hardware
70
60
50
40
Max (MB/s)
Min
30
Average
20
10
0
13
Worker
HP
Worker
HP
100MB
100MB
1GB
1GB
US ATLAS Tier2/Tier3 Workshop at
UChicago 2009-08-19
Conclusions




Xrootd can be configured to work with our
network constraints
Worker nodes need to be tuned for WAN
performance
Write performance is acceptable
Read performance is good
14
US ATLAS Tier2/Tier3 Workshop at
UChicago 2009-08-19
Proposed Production Configuration



Peered redirectors, one at each site, in a loadbalanced configuration. Provides fail-over if one
dies or if the link between the sites fails.
The link between the sites has never (yet) failed.
In addition, a supervisor node at each site.
Xrootd requires supervisors when the number
of data servers >64. MWT2 has 87.
15
US ATLAS Tier2/Tier3 Workshop at
UChicago 2009-08-19
Questions?
16
US ATLAS Tier2/Tier3 Workshop at
UChicago 2009-08-19