run-03

#!/bin/bash
set -eu

NODES=5
CORES_PER_NODE=8 # 16 VCPUs, but only 8 physical
CORES_PER_EXECUTOR=4
EXECUTORS_PER_NODE=$(( ${CORES_PER_NODE} / ${CORES_PER_EXECUTOR} ))
NUM_EXECUTORS=$(( ${NODES} * ${EXECUTORS_PER_NODE} ))
PARTITIONS=$(( ${NODES} * ${CORES_PER_NODE} * 2 ))

echo "Nodes: ${NODES}"
echo "Cores/node: ${CORES_PER_NODE}"
echo "Executors/node: ${EXECUTORS_PER_NODE}"
echo "Cores/executor: ${CORES_PER_EXECUTOR}"
echo "Executors: ${NUM_EXECUTORS}"
mkdir -p logs
sleep 5

for MULTIPLIER in `seq 5 10 130`; do
        N=$(( ${NODES} * ${CORES_PER_NODE} * ${MULTIPLIER} * 1000000 * 2))
        OUTPUT="logs/$(basename $0)-${N}.log"
	rm -f ${OUTPUT}
        echo "Number of trials: ${N}" | tee -a ${OUTPUT}
	echo "Data Size (MB): $(( ${N} * 8 / 1048576 ))" | tee -a ${OUTPUT}
        /usr/bin/time -f "TIMING: n=${N}, op=totalExecution, duration=%e" spark-submit --master yarn --deploy-mode client \
          --class com.cloudera.datascience.montecarlorisk.MonteCarloRisk \
          --conf "spark.dynamicAllocation.enabled=false" \
          --conf "spark.kryoserializer.buffer.max=512m" \
          --conf "spark.driver.maxResultSize=12576m" \
          --conf "spark.network.timeout=300s" \
          --driver-memory=40g \
          --executor-memory=9g \
          --num-executors=${NUM_EXECUTORS} \
          --executor-cores=${CORES_PER_EXECUTOR} \
          target/montecarlo-risk-0.0.1-SNAPSHOT.jar \
          data/instruments.csv ${N} ${PARTITIONS} data/means.csv data/covariances.csv \
          2>&1 | tee -a ${OUTPUT}
done