Difference between revisions of "CyberShake Computational Estimates"

From SCECpedia
Jump to navigationJump to search
 
(11 intermediate revisions by the same user not shown)
Line 1: Line 1:
 
We will describe or current best estimates for the CyberShake computational and data requirements as we progress in our simulation planning and testing. These estimates will help us identify which aspects of the CyberShake computational system needs to be optimized to work within our time and resource constraints.
 
We will describe or current best estimates for the CyberShake computational and data requirements as we progress in our simulation planning and testing. These estimates will help us identify which aspects of the CyberShake computational system needs to be optimized to work within our time and resource constraints.
  
== SGT Simulation Parameters ==
+
The UCERF 3 estimates assume that the number of ruptures increases from 15,000 to 350,000, but the number of rupture variations per rupture on average remains the same.
<pre>
 
1Hz CyberShake Estimates 2010.11.1
 
Max Freq 1Hz
 
Mesh Size Mesh Pts 12,000,000,000 320
 
Time Steps 40,000
 
 
Total Sites 4240
 
 
Total CPU Hours 320M
 
Estimated BW 10+Pflop/s
 
 
Estimated 1Hz Jaguar (half machine) 120 Days
 
Esimated 1Hz BW 40 Days
 
Estimated Effeciency Jagaur 8-10%
 
 
Jaguar 2.6GHz x 4FPU 10.4Gflop/s per core
 
Blue Waters 4GHz x 8 FPUs 32Gflop/s per core
 
  
</pre>
+
The 0.5 Hz numbers are taken from Study 14.2.
  
== Southern California, 0.5 Hz (current functionality) ==
+
The node-hours are estimates based on the XE6 and XK7 nodes on Blue Waters.
  
Sites: 223 sites (802 on 5 km grid)
+
== 1.0 Hz, 3 component ==
  
Jobs: 190 million
+
SGTs: At 0.5 Hz, it requires 38 GPU node-hrs per component.
 +
  (38 GPU node-hrs per component) x (3 components) x (8 times the gridpoints) x (2 times the timesteps) x (20% more efficient due to more work per GPU) = 1460 node-hrs per site.
  
CPU-hours: 5.5 million (Ranger)
+
23% of node hours to calculate SGTs.
  
Data products (seismograms, spectral acceleration): 2.1 TB
+
PP: At 0.5 Hz, it requires 41 CPU node-hrs per component.
 +
  (41 CPU node-hrs per components) x (3 components) x (25 times the rupture points) x (2 times the timesteps) x (20% more efficient due to rupture generator improvements) = 4920 node-hrs per site.
  
Runtime on half-Ranger: 174 hrs (7.3 days)
+
77% of node hours to calculate PP.
  
Runtime on half-Jaguar: 40 hrs (1.7 days)
+
Each site requires about 550,000 rupture variations (410,000 x 4/3 for rupture variations v3.3)
  
Runtime on half-BW(=half-Mira): 10 hrs
+
'''6380''' node-hours per 3-component site (181k core-hours)
  
Database entries: 366 million
+
'''1.82M''' node-hours for standard 3-component So Cal 286-site map (51.7M core-hours)
  
== Southern California, 1.0 Hz ==
+
'''5.73M''' node-hours for increased density 3-component So Cal 898-site map (162M core-hours)
  
AWP-ODC
+
'''8.93M''' node-hours for statewide adaptive 3-component California 1400-site map (253M core-hours)
  
SRFs increase by 25x to ~25 TB
+
== 2.0 Hz ==
  
Sites: 223 sites
+
SGTs: At 1.0 Hz, it requires 485 GPU node-hrs per component.
 +
  (485 GPU node-hrs per component) x (3 components) x (8 times the gridpoints) x (2 times the timesteps) = 23.3k node-hrs per site.
  
Jobs: 190 million
+
PP: At 1.0 Hz, it requires 1640 CPU node-hrs per component.
 +
  (1640 CPU node-hrs per components) x (3 components) x (2 times the timesteps) = 9.8k node-hrs per site.
  
CPU-hours: 19.3 million (Ranger)
+
'''33.1k''' node-hours per 3-component site (686k core-hours)
  
Data products (seismograms, spectral acceleration): 4.0 TB
+
'''9.47M''' node-hours for standard 3-component So Cal 286-site map (196M core-hours)
  
Runtime on half-Ranger: 613 hrs (25.5 days)
+
'''117M''' node-hours for increased density 3-component So Cal 3545-site map (2.4B core-hours)
  
Runtime on half-Jaguar: 142 hrs (5.9 days)
+
'''46.3M''' node-hours for statewide adaptive 3-component California 1400-site map (960M core-hours)
 
 
Runtime on half-BW(=half-Mira): 35 hrs (1.5 days)
 
 
 
Database entries: 366 million
 
 
 
== California, 0.5 Hz ==
 
 
 
=== Current software ===
 
 
 
Sites: 4240
 
 
 
Jobs: 3.6 billion
 
 
 
CPU-hours: 104.6 million (Ranger)
 
 
 
Data products (seismograms, spectral acceleration): 39.9 TB
 
 
 
Runtime on half-Ranger: 3322 hrs (138.4 days)
 
 
 
Runtime on half-Jaguar: 771 hrs (32.2 days)
 
 
 
Runtime on half-BW(=half-Mira): 192 hrs (8 days)
 
 
 
Database entries: 6.95 billion
 
 
 
=== With AWP-ODC ===
 
 
 
Sites: 4240
 
 
 
Jobs: 3.6 billion
 
 
 
CPU-hours: 83.5 million (Ranger)
 
 
 
Data products (seismograms, spectral acceleration): 39.9 TB
 
 
 
Runtime on half-Ranger: 2652 hrs (110.5 days)
 
 
 
Runtime on half-Jaguar: 616 hrs (25.7 days)
 
 
 
Runtime on half-BW(=half-Mira): 153 hrs (6.4 days)
 
 
 
Database entries: 6.95 billion
 
 
 
== California, 1.0 Hz ==
 
 
 
AWP-ODC
 
 
 
SRFs increase by 25x to ~25 TB
 
 
 
[http://hypocenter.usc.edu/research/cybershake/CA_10km_sites.png Site Map: 4240]
 
 
 
Jobs: 3.6 billion
 
 
 
CPU-hours: 376.2 million (337.8 million - SGT generation only, 339.1 million - SGT workflow)
 
 
 
Data products (seismograms, spectral acceleration): 76.5 TB
 
 
 
Runtime on half-Ranger: 11947 hrs (497.8 days)
 
 
 
Runtime on half-Jaguar: 3096 hrs (129 days)
 
 
 
Runtime on half-BW(=half-Mira): 770 hrs (32.1 days) (4.3% of yearly CPU-hrs)
 
 
 
Database entries: 6.95 billion
 

Latest revision as of 15:52, 20 June 2014

We will describe or current best estimates for the CyberShake computational and data requirements as we progress in our simulation planning and testing. These estimates will help us identify which aspects of the CyberShake computational system needs to be optimized to work within our time and resource constraints.

The UCERF 3 estimates assume that the number of ruptures increases from 15,000 to 350,000, but the number of rupture variations per rupture on average remains the same.

The 0.5 Hz numbers are taken from Study 14.2.

The node-hours are estimates based on the XE6 and XK7 nodes on Blue Waters.

1.0 Hz, 3 component

SGTs: At 0.5 Hz, it requires 38 GPU node-hrs per component.

 (38 GPU node-hrs per component) x (3 components) x (8 times the gridpoints) x (2 times the timesteps) x (20% more efficient due to more work per GPU) = 1460 node-hrs per site.

23% of node hours to calculate SGTs.

PP: At 0.5 Hz, it requires 41 CPU node-hrs per component.

 (41 CPU node-hrs per components) x (3 components) x (25 times the rupture points) x (2 times the timesteps) x (20% more efficient due to rupture generator improvements) = 4920 node-hrs per site.

77% of node hours to calculate PP.

Each site requires about 550,000 rupture variations (410,000 x 4/3 for rupture variations v3.3)

6380 node-hours per 3-component site (181k core-hours)

1.82M node-hours for standard 3-component So Cal 286-site map (51.7M core-hours)

5.73M node-hours for increased density 3-component So Cal 898-site map (162M core-hours)

8.93M node-hours for statewide adaptive 3-component California 1400-site map (253M core-hours)

2.0 Hz

SGTs: At 1.0 Hz, it requires 485 GPU node-hrs per component.

 (485 GPU node-hrs per component) x (3 components) x (8 times the gridpoints) x (2 times the timesteps) = 23.3k node-hrs per site.

PP: At 1.0 Hz, it requires 1640 CPU node-hrs per component.

 (1640 CPU node-hrs per components) x (3 components) x (2 times the timesteps) = 9.8k node-hrs per site.

33.1k node-hours per 3-component site (686k core-hours)

9.47M node-hours for standard 3-component So Cal 286-site map (196M core-hours)

117M node-hours for increased density 3-component So Cal 3545-site map (2.4B core-hours)

46.3M node-hours for statewide adaptive 3-component California 1400-site map (960M core-hours)