CSEP - ETAS Simulation Plan

This page summarizes the performance study ran on Stampede 3 in order to calculate the simulation requirements for computing 1 day ETAS forecasts from 2007 to 2018.

Stampede 3 Installation and Configuration

• Java - jdk-21.0.1+12
• FastMPJ

The Slurm file header used to run the simulations is shown below:

#SBATCH -t 6:00:00
#SBATCH -N 7
#SBATCH -n 336
#SBATCH -p spr
#SBATCH -A DS-Cybershake

where the "-n" parameter is calculated as 48 * number of nodes (N). Additionally, the following parameters were used in the Slurm script

MEM_GIGS=110
THREADS=20
FMPJ_HOME=/work2/02404/fsilva/stampede3/FastMPJ
#CLEAN_OPTION="--clean"
export FMPJ_HOME=/work2/02404/fsilva/stampede3/FastMPJ

Performance Results

To measure the Stampede 3 performance running the one day ETAS forecasts, we ran the same simulation scenario using different numbers of nodes. To compute these results, we used the same random seed and date. We also deleted the cached results stored in the scratch filesystem after each simulation in order to force the recalculation of the entire run. The command-line used to generate the run was:

u3etas_comcat_config_builder.sh --end-time 1717484400000 --num-simulations 100000 --duration-years 0.002737851 --include-spontaneous --historical-catalog --start-after-historical --etas-k-cov 1.5 --random-seed 123456789 --hpc-site TACC_FRONTERA --nodes 35 --hours 24 --queue normal --output-dir $ETAS_SIM_DIR/2024_06_04-ComcatPlusHistorical-Start20240604_1day_100000Simulations_Statewide_PointSources_kCOV1p5_Spontaneous_HistCatalog --binary-output

The results are as follows:

Number of Nodes	Runtime (min)	Service Units (SUs) Used
7	169	19.7
14	84	19.6
28	43	20.1
56	24	22.4

In the table above, service units was computed by dividing the runtime by 60 and multiplying by the number of nodes used. We observe that as the numbers of nodes used increased, runtime decreased linearly while the allocation usage remained mostly flat.

Requirements for the Complete ETAS Simulation Runs

Using the performance results obtained above, we can calculate the requirements for the full simulation run by multiplying the required SUs for a single run by the total number of runs:

• Start Date = 1 August 2007
• End Date = 30 August 2018
• Total Number of Days = 4045

For the total required service units, we multiple the total number of 1 day forecasts (4045) by the number of service units used for each run (20):

• SUs needed = 20 * 4045 = 80900 SUs

The storage requirements per 1 day forecast is as follows:

• Binary results (results_*.bin files) ~ 100M
• Complete output folder (with logs) ~ 272M

If we multiply the numbers above by the total number of 1-day forecasts, we have

• Total storage for data = 100M * 4045 = 405G
• Total storage (including logs) = 272M * 4045 = 1.1T

ETAS Production Run Setup

For the UCERF3-ETAS production run setup on Stampede 3, we split the simulation period into 134 30-day bundles using UTC times. We also used the following parameters for each bundle:

• Number of nodes: 40
• Node type: Skylake
• Wall time: 24 hours
• Memory: 110 GB
• Threads: 20
• Options: TEMP_OPTION, SCRATCH_OPTION

Slurm header:

#SBATCH -t 24:00:00
#SBATCH -N 40
#SBATCH -n 1920
#SBATCH -p skx

ETAS Production Run Status

Period
12/2006-12/2007								000	001	002	003	004
12/2007-12/2008	005	006	007	008	009	010	011	012	013	014	015	016
12/2008-12/2009	017	018	019	020	021	022	023	024	025	026	027	028
12/2009-12/2010	029	030	031	032	033	034	035	036	037	038	039	040
12/2010-12/2011	041	042	043	044	045	046	047	048	049	050	051	052
12/2011-12/2012	053	054	055	056	057	058	059	060	061	062	063	064
12/2012-12/2013	065	066	067	068	069	070	071	072	073	074	075	076
12/2013-12/2014	077	078	079	080	081	082	083	084	085	086	087	088
12/2014-12/2015	089	090	091	092	093	094	095	096	097	098	099	100
12/2015-12/2016	101	102	103	104	105	106	107	108	109	110	111	112
12/2016-12/2017	113	114	115	116	117	118	119	120	121	122	123	124
12/2017-12/2018	125	126	127	128	129	130	131	132	133

Completed

Running

Queued