Real Life Stories on Extreme Performance with In-Memory Database Technology Presented at Oracle Open World
Real Life Stories on Extreme Performance with
In-Memory Database Technology
Presented at Oracle Open World
The Lockheed Martin Census Practice
TimesTen Use in UK Census
John White
Chief Architect – UK Census
IS&GS-Civil, Greenbelt, MD
OpenWorld 2011
© 2011 Lockheed Martin Corporation. All Rights Reserved
Lockheed Martin – We Are the How
Between the Idea and the Achievement, There Is One Important Word:
HOW … And It Is the HOW that Makes All the Difference
© 2011 Lockheed Martin Corporation. All Rights Reserved
Use Case: UK 2011 Census
$-
$5.00
$10.00
$15.00
$20.00
$25.00
17
90
18
10
18
30
18
50
18
70
18
90
19
10
19
30
19
50
19
70
19
90
Normalized Response Cost over time
Cost by Response Channel •Field - ~$90
•Telephone - ~$10
•Paper - ~$4
•Internet - ~$1
© 2011 Lockheed Martin Corporation. All Rights Reserved
Challenge: Central DB Load
• Requirement: Must store data after each page
• Initial IDB Load Estimate: 465 TPS
• Final IDB Load Estimate: 119,026 TPS
Change Increased load on IDB
Load Balancer design solution 2X
WSH integration/login redirection solution 2X
24->32 page questionnaires 1.6X
Questionnaire App Database Interaction 8X
Increased Internet Uptake 5X
Total 2 * 2 * 1.6 * 8 * 5 = 256X
© 2011 Lockheed Martin Corporation. All Rights Reserved
Use of Times 10
Rearchitected solution to reduce load on IDB
Used TimesTen as local data cache in cluster
Served local transactions
Read from IDB during login
Aged to IDB after local transaction
Enabled move from risky Active/Active
to Active/Passive IDB
Concerns
Deployment Location
Sizing
Aging latency
Network load
IDB
App Svr
BEFORE
AFTER
RDBMS, RAC
…
App Svr …
IDB
RDBMS, RAC
XDB
… TimesTen
© 2011 Lockheed Martin Corporation. All Rights Reserved
WAN (35Mbps)
SIG (P)
FW
LB
LB
FW
IDC
AS
A
S
AS
Internet
SIG (P)
FW
LB
LB
FW
WS
W
S
IDC
LB
WS
W
S
LB
WS
W
S
WS
WS
WS
WS
SIG (P) SIG (P)
WS
Before
AS
WSH WSH
WS
WS
W
S
WS
WS
A
S
AS
AS
A
S
AS
A
S
AS
A
S
FW FW
Advanced Replication
UDBC
FW FW
IDB
IDB
RDBMS
UDBC
RDBMS
RDBMS, RAC
IDB
IDB
RDBMS, RAC
© 2011 Lockheed Martin Corporation. All Rights Reserved
WAN (65Mbps)
SIG (P)
FW
LB
WS
W
S
IDC
AS
A
S
AS
A
S
Internet
SIG (P)
FW
LB
FW
WS
W
S
IDC
AS
A
S
AS
A
S
FW FW
SIG (S)
FW
LB
WS
W
S
IDC
AS
A
S
AS
A
S
SIG (S)
FW
LB
WS
W
S
IDC A
S
AS
AS
A
S
LB
WS
W
S
LB
WS
WS
WS
WS
WS
WS
After
WSH WSH
FW FW
WS
X1
0
X1
0
X1
0
X1
0
X1
0
X1
0
X1
0
X1
0
WS
WS
W
S
WS
WS
W
S
WS
FW
DataGuard
TimesTen
UDBC
UDBC
UDBC
UDBC
RDBMS, RAC RDBMS, RAC
RDBMS, RAC
IDB
IDB
RDBMS, RAC
TimesTen TimesTen
IDB
IDB
© 2011 Lockheed Martin Corporation. All Rights Reserved
Results
Verified with cloud based full scale load testing:
3.0 million daily responses
200,000 peak concurrent users
1 second average page response time
IDB: 6,118 TPS
Availability maintained through simulated site, cluster and
component failure
Actual production usage:
3.7 million total responses
36,000 peak concurrent users
<1 second average page response time
IDB: 1,101 TPS
Production availability: 99.94%
© 2011 Lockheed Martin Corporation. All Rights Reserved
Lessons Learned
Performance better than expected – 10X better
Needed more RAM than expected – 24GB/CPU
Log writing greater than expected
Data written to XDB disk then aged to central IDB
Disks need to be sized to handle transaction load
TimesTen deployment doesn’t have to be coresident with app
Latency somewhat configurable
Max Latency = aging frequency + max time to age
Network load not overly burdensome
On par with Advanced Replication traffic load - ~10Mbps
Cloud based testing was most cost effective to do full load test