Upload
others
View
5
Download
0
Embed Size (px)
Citation preview
Experiences Using Windows Azure
to Calibrate Watershed Models
Marty Humphrey, Norm Beekwilder University of Virginia
Jon Goodall, Mehmet Ercan
University of South Carolina
Mississippi River Watershed Chesapeake Bay Watershed
Example Large Watersheds
Typical Scale of Watershed Models
Eno River near Durham, NC
Scale
• Mississippi: 1,245,000 sq mi (3,220,000 km2)
• Chesapeake: 64,000 sq mi (166,000 km2)
• Eno (our case study watershed): 66 sq mi (171 km2)
Eno to Chesapeake (~ 1,000 times)
Eno to Mississippi (~ 20,000 times)
Watershed Hydrology
SWAT
Model
Source: SWAT Theoretical Document
Challenges in Watershed Modeling
• Data Preparation
– Data exists, but files are large and require preprocessing
• Model Calibration
– Requires running the model multiple times with varying parameters
• Scale up Model to Large System (Chesapeake, Mississippi)
– Impractical using current approaches
HPC Cluster
Azure Compute Instances
Azure Compute Proxies
Early Results
(Cloud Futures, June 2011)
Stage-in Compute Total
Scientist laptop 0 55 sec 55 sec
Win2 5 sec 60 sec 65 sec
Azure (ex-large) 53 sec 32 sec 85 sec
Issues
(Cloud Futures, June 2011)
• Windows Azure only or cloudbursting?
• Data storage – where, how?
• Data sharing/reuse policy?
• Task granularity / coding?
• Task synchronization (e.g., MPI)?
# Parameters
# USGS gage ID
# Other settings
SWAT
TxtInOut
E and parameter values
SWAT
TxtInOut
Parameters
& Settings
Get Streamflow
USGS gage ID
E values
Analyze E Values
and Create New
Parameter Series
The Calibration
Method
Is the calibration
criteria satisfied?
Stop
Yes
No
Parameter
Series
CUASHI
Streamflow
Edit Input Files
Run SWAT
Calculate E
SWAT
TxtInOut
Results
P-DDS
Choose parameters/values
Execute
model
Execute
model
Execute
model
Execute
model
Change candidate parameters/distributions
Choose Best Value
P-DDS
0
2
4
6
8
10
12
14
16
18
1
15
9
31
7
47
5
63
3
79
1
94
9
11
07
12
65
14
23
15
81
17
39
18
97
20
55
22
13
23
71
25
29
26
87
28
45
30
03
31
61
33
19
34
77
36
35
37
93
39
51
41
09
42
67
44
25
45
83
47
41
48
99
50
57
52
15
53
73
55
31
56
89
58
47
60
05
61
63
63
21
64
79
66
37
67
95
69
53
71
11
72
69
74
27
75
85
77
43
79
01
80
59
82
17
83
75
85
33
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
1
24
47
70
93
11
6
13
9
16
2
18
5
20
8
23
1
25
4
27
7
30
0
32
3
34
6
36
9
39
2
41
5
43
8
46
1
48
4
50
7
53
0
55
3
57
6
59
9
62
2
64
5
66
8
69
1
71
4
73
7
76
0
78
3
80
6
82
9
85
2
87
5
89
8
92
1
94
4
96
7
99
0
Verification
Stage Out
Compute
Stage In
NW-DDS
Choose parameters/values
Execute
model
Execute
model
Execute
model
Execute
model
Change candidate parameters/distributions
NW-DDS
0
2
4
6
8
10
12
14
16
18
1
48
95
14
2
18
9
23
6
28
3
33
0
37
7
42
4
47
1
51
8
56
5
61
2
65
9
70
6
75
3
80
0
84
7
89
4
94
1
98
8
10
35
10
82
11
29
11
76
12
23
12
70
13
17
13
64
14
11
14
58
15
05
15
52
15
99
16
46
16
93
17
40
17
87
18
34
18
81
19
28
19
75
20
22
20
69
21
16
21
63
22
10
22
57
23
04
23
51
23
98
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
1
24
47
70
93
11
6
13
9
16
2
18
5
20
8
23
1
25
4
27
7
30
0
32
3
34
6
36
9
39
2
41
5
43
8
46
1
48
4
50
7
53
0
55
3
57
6
59
9
62
2
64
5
66
8
69
1
71
4
73
7
76
0
78
3
80
6
82
9
85
2
87
5
89
8
92
1
94
4
96
7
99
0
Verification
Stage Out
Compute
Stage In
0
2
4
6
8
10
12
14
161
51
10
1
15
1
20
1
25
1
30
1
35
1
40
1
45
1
50
1
55
1
60
1
65
1
70
1
75
1
80
1
85
1
90
1
95
1
10
01
10
51
11
01
11
51
12
01
12
51
13
01
13
51
14
01
14
51
15
01
15
51
16
01
16
51
17
01
17
51
18
01
18
51
19
01
19
51
20
01
20
51
21
01
21
51
22
01
22
51
23
01
23
51
24
01
24
51
25
01
25
51
0
10
20
30
40
50
60
1
15
29
43
57
71
85
99
11
3
12
7
14
1
15
5
16
9
18
3
19
7
21
1
22
5
23
9
25
3
26
7
28
1
29
5
30
9
32
3
33
7
35
1
36
5
37
9
39
3
40
7
42
1
43
5
44
9
46
3
47
7
49
1
50
5
51
9
53
3
54
7
56
1
57
5
58
9
60
3
61
7
63
1
64
5
65
9
67
3
68
7
70
1
0
50
100
150
200
2501 7
13
19
25
31
37
43
49
55
61
67
73
79
85
91
97
10
3
10
9
11
5
12
1
12
7
13
3
13
9
14
5
15
1
15
7
16
3
16
9
17
5
18
1
18
7
19
3
19
9
20
5
21
1
21
7
22
3
22
9
23
5
24
1
24
7
25
3
25
9
26
5
27
1
27
7
28
3
28
9
29
5
30
1
Efficiency
Duration Speedup
Desktop, SWAT internal
calibration (single
threaded)
11.4 hours (41040 sec) --
Windows Azure, Ex-large
VM, 16 cores, NW-DDS
43.32 min (2599 sec) 15.78x
Windows Azure, Ex-large
VM, 64 cores, NW-DDS
11.76 min (706 sec) 58.13x
Windows Azure, Ex-large
VM, 256 cores, NW-DDS
5.03 min (302 sec) 135.89x
As number of cores increase, dominated by [a] stragglers, and
[b] re-running best param set to get output files
Re-visiting the Issues
(Cloud Futures, June 2011)
• Windows Azure only or cloudbursting?
– Only Windows Azure (with head node inside enterprise)
• Data storage – where, how?
– Only Windows Azure, so [a] “where” is largely non-issue, and [b] “how” is as little as possible
• Data sharing/reuse policy?
– Not as much as an issue as we would like
• Task granularity / coding?
– Hmm… not exactly an issue (see next slide)
• Task synchronization (e.g., MPI)?
– Removed in an uninteresting way
Next steps: Calibration
• Additional watershed models (e.g., HSPF)
• Additional calibration algorithms
• Better control of search (e.g., stopping criteria)
– Visual results presentation and visual steering
– Trading off fast-vs.-more-exhaustive
• Better sharing / non-sharing
– Insta-virtual-cluster a la HadoopOnAzure
• How does insta-calibration change the watershed model design process? (continual calibration?)
Next steps: Data preparation
Observations on Windows Azure
• Does Windows Azure (more broadly: cloud computing) substantially change the challenges of collaboration between domains? No.
• What do we like?
– Cloudbursting mechanism (head node, VPN).
– Predictability is sufficient (and continues to improve).
• What would we like to see?
– Faster boot time for VMs.
– Non-round-up cost-model.
– Faster speed of light.
Summary
• Goals: Watershed calibration, data prep, and
large-scale modeling
• Good progress on watershed calibration, data
prep next
• Basic issues of collaboration across domains
will continue to dominate