Difference between revisions of "UCERF3 JetStream Benchmark"

From SCECpedia
Jump to navigationJump to search
(Created page with " == Computing Proposal == The Collaboratory for the Study of Earthquake Predictability (CSEP) provides a computational infrastructure to conduct numerical experiments that ev...")
 
 
(3 intermediate revisions by the same user not shown)
Line 1: Line 1:
 +
== Introduction ==
 +
JetStream is accessed through the [https://use.jetstream-cloud.org/application/dashboard Atmosphere] web portal. Read the [https://portal.xsede.org/jetstream#qsguide:launch user documentation] to launch a virtual machine on the JetStream system using Atmosphere.
  
== Computing Proposal ==
+
To connect via SSH, configure an SSH key and upload this in the ‘advanced’ section of the JetStream user settings. In addition, scripts can be uploaded to set up an instance that will configure the environment of the instance.
 +
 
 +
 
 +
== Benchmark Results ==
 +
{| class = 'wikitable'
 +
|-
 +
! SIMULATION
 +
! HORIZON
 +
! NUM SIMULATIONS
 +
! THREADS
 +
! SIM / THREAD
 +
! RUNTIME (s)
 +
! MEMORY USAGE (GB)
 +
! TIME / SIMULATION
 +
! TIME / THREAD
 +
! SUs
 +
! MACHINE
 +
|-
 +
| 1
 +
| 1
 +
| 10
 +
| 5
 +
| 2
 +
| 477.2749898
 +
| 25
 +
| 47.72749898
 +
| 95.45499797
 +
|
 +
| m1.large
 +
|-
 +
| 2
 +
| 1
 +
| 25
 +
| 5
 +
| 5
 +
| 1015.450814
 +
| 25
 +
| 40.61803255
 +
| 203.0901628
 +
|
 +
| m1.large
 +
|-
 +
| 3
 +
| 1
 +
| 50
 +
| 10
 +
| 5
 +
| 1879.826009
 +
| 50
 +
| 37.59652019
 +
| 187.9826009
 +
|
 +
| m1.xxlarge
 +
|-
 +
| 4
 +
| 1
 +
| 100
 +
| 10
 +
| 10
 +
| 3535.324325
 +
| 50
 +
| 35.35324325
 +
| 353.5324325
 +
|
 +
| m1.xxlarge
 +
|-
 +
| 5
 +
| 1
 +
| 200
 +
| 10
 +
| 20
 +
| 9787.439788
 +
| 50
 +
| 48.93719894
 +
| 978.7439788
 +
|
 +
| m1.xxlarge
 +
|-
 +
| 6
 +
| 1
 +
| 200
 +
| 20
 +
| 10
 +
| 12505.12893
 +
| 100
 +
| 62.52564464
 +
| 625.2564464
 +
|
 +
| m1.xxlarge
 +
|-
 +
| 7
 +
| 0.08333333
 +
| 500
 +
| 20
 +
| 25
 +
| 27366.20519
 +
| 100
 +
| 54.73241037
 +
| 1368.310259
 +
|
 +
| m1.xxlarge
 +
|-
 +
| 8
 +
| 1
 +
| 500
 +
| 20
 +
| 25
 +
| 30934.89149
 +
| 100
 +
| 61.86978297
 +
| 1546.744574
 +
|
 +
| m1.xxlarge
 +
|-
 +
| 9
 +
| 1
 +
| 20
 +
| 20
 +
| 1
 +
| 1175.592247
 +
| 100
 +
| 58.77961237
 +
| 58.77961237
 +
|
 +
| m1.xxlarge
 +
|-
 +
| 10
 +
| 1
 +
| 10
 +
| 10
 +
| 1
 +
| 470.5639017
 +
| 50
 +
| 47.05639017
 +
| 47.05639017
 +
|
 +
| m1.xxlarge
 +
|-
 +
| 11
 +
| 1
 +
| 5
 +
| 5
 +
| 1
 +
| 240.4323306
 +
| 25
 +
| 48.08646612
 +
| 48.08646612
 +
|
 +
| m1.xxlarge
 +
|-
 +
| 12
 +
| 1
 +
| 1
 +
| 1
 +
| 1
 +
| 117.9551687
 +
| 5
 +
| 117.9551687
 +
| 117.9551687
 +
|
 +
| m1.xxlarge
 +
|}
 +
 
 +
== Computing Proposal for JetStream ==
  
 
The Collaboratory for the Study of Earthquake Predictability (CSEP) provides a computational infrastructure to conduct numerical experiments that evaluate earthquake forecasts against authoritative data sources. Experiments can be either retrospective or prospective with each having different computational requirements. Prospective experiments involve calculations that trigger automatically from recent seismicity or on regular intervals (e.g., daily, weekly, monthly, etc). Retrospective experiments are initiated on-demand by scientists to evaluate earthquake forecasts over periods with recorded earthquake data. We are developing a software infrastructure that will support both type of experiments and provide thoughtful data access to the CSEP community. Additionally, we plan to explore the possibility of developing the CSEP infrastructure into a SaaS that gives researchers the ability to conduct bespoke experiments in a controlled and secure environment.
 
The Collaboratory for the Study of Earthquake Predictability (CSEP) provides a computational infrastructure to conduct numerical experiments that evaluate earthquake forecasts against authoritative data sources. Experiments can be either retrospective or prospective with each having different computational requirements. Prospective experiments involve calculations that trigger automatically from recent seismicity or on regular intervals (e.g., daily, weekly, monthly, etc). Retrospective experiments are initiated on-demand by scientists to evaluate earthquake forecasts over periods with recorded earthquake data. We are developing a software infrastructure that will support both type of experiments and provide thoughtful data access to the CSEP community. Additionally, we plan to explore the possibility of developing the CSEP infrastructure into a SaaS that gives researchers the ability to conduct bespoke experiments in a controlled and secure environment.
  
Our initial CSEP experiment will focus on the Uniform California Earthquake Rupture Forecast Version 3 (UCERF3-ETAS) [Field et al., 2017]. The standard unit-of-work for this UCERF3-ETAS experiment is a one-year stochastic event sets consisting of 10,000 synthetic catalogs. We have benchmarked the unit-of-work to use about 8000 SUs on JetStream and about 2.5 Gb of generated data. We will address our understanding of the added benefits that finite-faults provide in earthquake forecasts over hypocenter based models. In addition, we will evaluate how UCERF3-ETAS performs relative to more simple forecasts for all earthquakes with Mw > 6.2 that have occured since 1985. To accomplish these goals we estimate the runs will require about 25 units-of-work or 200,000 SUs. We require ten (m1.xxlarge) instances to perform these calculations with a shared mounted 100 Gb volume. We will also need 1 Tb of persistent storage to store metadata and archived experiment data. We would also like to request API access to OpenStack so we can automatically provision and free resources as necessary.  
+
Our initial CSEP experiment will focus on the Uniform California Earthquake Rupture Forecast Version 3 (UCERF3-ETAS) [Field et al., 2017]. The standard unit-of-work for this UCERF3-ETAS experiment is a one-year stochastic event sets consisting of 10,000 synthetic catalogs. We have benchmarked the unit-of-work to use about 8000 SUs on JetStream and about 2.5 Gb of generated data. We will address our understanding of the added benefits that finite-faults provide in earthquake forecasts over hypocenter based models. In addition, we will evaluate how UCERF3-ETAS performs relative to more simple forecasts for all earthquakes with Mw > 6.2 that have occured since 1985. To accomplish these goals we estimate the runs will require about 25 units-of-work or 200,000 SUs. We require ten (m1.xxlarge) instances to perform these calculations with a shared mounted 100 Gb volume. We will also need 1 Tb of persistent storage to store metadata and archived experiment data. We would also like to request API access to OpenStack so we can automatically provision and free resources as necessary.
 
 
  
 
== References ==
 
== References ==
  
 
Field, Edward H., et al. "A spatiotemporal clustering model for the Third Uniform California Earthquake Rupture Forecast (UCERF3‐ETAS): Toward an operational earthquake forecast." Bulletin of the Seismological Society of America 107.3 (2017): 1049-1081.
 
Field, Edward H., et al. "A spatiotemporal clustering model for the Third Uniform California Earthquake Rupture Forecast (UCERF3‐ETAS): Toward an operational earthquake forecast." Bulletin of the Seismological Society of America 107.3 (2017): 1049-1081.

Latest revision as of 19:50, 26 October 2018

Introduction

JetStream is accessed through the Atmosphere web portal. Read the user documentation to launch a virtual machine on the JetStream system using Atmosphere.

To connect via SSH, configure an SSH key and upload this in the ‘advanced’ section of the JetStream user settings. In addition, scripts can be uploaded to set up an instance that will configure the environment of the instance.


Benchmark Results

SIMULATION HORIZON NUM SIMULATIONS THREADS SIM / THREAD RUNTIME (s) MEMORY USAGE (GB) TIME / SIMULATION TIME / THREAD SUs MACHINE
1 1 10 5 2 477.2749898 25 47.72749898 95.45499797 m1.large
2 1 25 5 5 1015.450814 25 40.61803255 203.0901628 m1.large
3 1 50 10 5 1879.826009 50 37.59652019 187.9826009 m1.xxlarge
4 1 100 10 10 3535.324325 50 35.35324325 353.5324325 m1.xxlarge
5 1 200 10 20 9787.439788 50 48.93719894 978.7439788 m1.xxlarge
6 1 200 20 10 12505.12893 100 62.52564464 625.2564464 m1.xxlarge
7 0.08333333 500 20 25 27366.20519 100 54.73241037 1368.310259 m1.xxlarge
8 1 500 20 25 30934.89149 100 61.86978297 1546.744574 m1.xxlarge
9 1 20 20 1 1175.592247 100 58.77961237 58.77961237 m1.xxlarge
10 1 10 10 1 470.5639017 50 47.05639017 47.05639017 m1.xxlarge
11 1 5 5 1 240.4323306 25 48.08646612 48.08646612 m1.xxlarge
12 1 1 1 1 117.9551687 5 117.9551687 117.9551687 m1.xxlarge

Computing Proposal for JetStream

The Collaboratory for the Study of Earthquake Predictability (CSEP) provides a computational infrastructure to conduct numerical experiments that evaluate earthquake forecasts against authoritative data sources. Experiments can be either retrospective or prospective with each having different computational requirements. Prospective experiments involve calculations that trigger automatically from recent seismicity or on regular intervals (e.g., daily, weekly, monthly, etc). Retrospective experiments are initiated on-demand by scientists to evaluate earthquake forecasts over periods with recorded earthquake data. We are developing a software infrastructure that will support both type of experiments and provide thoughtful data access to the CSEP community. Additionally, we plan to explore the possibility of developing the CSEP infrastructure into a SaaS that gives researchers the ability to conduct bespoke experiments in a controlled and secure environment.

Our initial CSEP experiment will focus on the Uniform California Earthquake Rupture Forecast Version 3 (UCERF3-ETAS) [Field et al., 2017]. The standard unit-of-work for this UCERF3-ETAS experiment is a one-year stochastic event sets consisting of 10,000 synthetic catalogs. We have benchmarked the unit-of-work to use about 8000 SUs on JetStream and about 2.5 Gb of generated data. We will address our understanding of the added benefits that finite-faults provide in earthquake forecasts over hypocenter based models. In addition, we will evaluate how UCERF3-ETAS performs relative to more simple forecasts for all earthquakes with Mw > 6.2 that have occured since 1985. To accomplish these goals we estimate the runs will require about 25 units-of-work or 200,000 SUs. We require ten (m1.xxlarge) instances to perform these calculations with a shared mounted 100 Gb volume. We will also need 1 Tb of persistent storage to store metadata and archived experiment data. We would also like to request API access to OpenStack so we can automatically provision and free resources as necessary.

References

Field, Edward H., et al. "A spatiotemporal clustering model for the Third Uniform California Earthquake Rupture Forecast (UCERF3‐ETAS): Toward an operational earthquake forecast." Bulletin of the Seismological Society of America 107.3 (2017): 1049-1081.