Commit b1d3b47b authored by Pawel Niedzwiecki's avatar Pawel Niedzwiecki Committed by Jim Harris
Browse files

test/nvme: Test case 1 and 2 of NVME performance tests 



Measure Iops/core performance

Change-Id: I5c2632b45c62cadbd46a67a7284ece9f77b70902
Signed-off-by: default avatarPawel Niedzwiecki <pawelx.niedzwiecki@intel.com>
Reviewed-on: https://review.gerrithub.io/c/spdk/spdk/+/423170


Tested-by: default avatarSPDK CI Jenkins <sys_sgci@intel.com>
Reviewed-by: default avatarJohn Kariuki <John.K.Kariuki@intel.com>
Reviewed-by: default avatarKarol Latecki <karol.latecki@intel.com>
Reviewed-by: default avatarJim Harris <james.r.harris@intel.com>
Reviewed-by: default avatarBen Walker <benjamin.walker@intel.com>
parent 933270ea

test/nvme/perf/common.sh

0 → 100755
+266 −0
Original line number Diff line number Diff line 
#!/usr/bin/env bash



set -xe

BASE_DIR=$(readlink -f $(dirname $0))

ROOT_DIR=$(readlink -f $BASE_DIR/../../..)

PLUGIN_DIR_NVME=$ROOT_DIR/examples/nvme/fio_plugin

PLUGIN_DIR_BDEV=$ROOT_DIR/examples/bdev/fio_plugin

. $ROOT_DIR/scripts/common.sh || exit 1

. $ROOT_DIR/test/common/autotest_common.sh

NVME_FIO_RESULTS=$BASE_DIR/result.json



PRECONDITIONING=true

FIO_BIN="/usr/src/fio/fio"

RUNTIME=600

PLUGIN="nvme"

RAMP_TIME=30

BLK_SIZE=4096

RW=randrw

MIX=100

TYPE=("randread" "randrw" "randwrite")

IODEPTH=256

DISKNO=1

ONEWORKLOAD=false

CPUS_ALLOWED=1

REPEAT_NO=3



function get_cores(){

	local cpu_list="$1"

	for cpu in ${cpu_list//,/ }; do

		echo $cpu

	done

}



function get_cores_numa_node(){

	local cores=$1

	for core in $cores; do

		echo $(lscpu -p=cpu,node | grep "^$core\b" | awk -F ',' '{print $2}')

	done

}



function get_numa_node(){

	local plugin=$1

	local disks=$2

	if [ "$plugin" = "nvme" ]; then

		for bdf in $disks; do

			local driver=`grep DRIVER /sys/bus/pci/devices/$bdf/uevent |awk -F"=" '{print $2}'`

			# Use this check to ommit blacklisted devices ( not binded to driver with setup.sh script )

			if [ "$driver" = "vfio-pci" ] || [ "$driver" = "uio_pci_generic" ]; then

				echo $(cat /sys/bus/pci/devices/$bdf/numa_node)

			fi

		done

	elif [ "$plugin" = "bdev" ]; then

		local bdevs=$(discover_bdevs $ROOT_DIR $BASE_DIR/bdev.conf)

		for name in $disks; do

			local bdev_bdf=$(jq -r ".[] | select(.name==\"$name\").driver_specific.nvme.pci_address" <<< $bdevs)

			echo $(cat /sys/bus/pci/devices/$bdev_bdf/numa_node)

		done

	fi

}



function get_disks(){

	local plugin=$1

	if [ "$plugin" = "nvme" ]; then

		for bdf in $(iter_pci_class_code 01 08 02); do

			driver=`grep DRIVER /sys/bus/pci/devices/$bdf/uevent |awk -F"=" '{print $2}'`

			if [ "$driver" = "vfio-pci" ] || [ "$driver" = "uio_pci_generic" ]; then

				echo "$bdf"

			fi

		done

	elif [ "$plugin" = "bdev" ]; then

		local bdevs=$(discover_bdevs $ROOT_DIR $BASE_DIR/bdev.conf)

		echo $(jq -r '.[].name' <<< $bdevs)

	fi

}



function get_disks_on_numa(){

	local devs=($1)

	local numas=($2)

	local numa_no=$3

	local disks_on_numa=""

	local i



	for (( i=0; i<${#devs[@]}; i++ ))

	do

		if [ ${numas[$i]} = $numa_no ]; then

			disks_on_numa=$(($disks_on_numa+1))

		fi

	done

	echo $disks_on_numa

}



function create_fio_config(){

	local disk_no=$1

	local plugin=$2

	local disks=($3)

	local disks_numa=($4)

	local cores=($5)

	local total_disks=${#disks[@]}

	local no_cores=${#cores[@]}

	local filename=""



	local cores_numa=($(get_cores_numa_node "$5"))

	local disks_per_core=$(($disk_no/$no_cores))

	local disks_per_core_mod=$(($disk_no%$no_cores))

	for (( i=0; i<$no_cores; i++ ))

	do

		core_numa=${cores_numa[$i]}

		total_disks_per_core=$disks_per_core

		if [ "$disks_per_core_mod" -gt "0" ]; then

			total_disks_per_core=$(($disks_per_core+1))

			disks_per_core_mod=$(($disks_per_core_mod-1))

		fi



		if [ "$total_disks_per_core" = "0" ]; then

			break

		fi



		sed -i -e "\$a[filename${i}]" $BASE_DIR/config.fio

		#use cpus_allowed as cpumask works only for cores 1-32

		sed -i -e "\$acpus_allowed=${cores[$i]}" $BASE_DIR/config.fio

		m=0	#counter of disks per cpu core numa

		n=0 #counter of all disks

		while [ "$m" -lt  "$total_disks_per_core" ]; do

			if [ "${disks_numa[$n]}" = "$core_numa" ]; then

				m=$(($m+1))

				if [ "$plugin" = "nvme" ]; then

					filename='trtype=PCIe traddr='${disks[$n]//:/.}' ns=1'

				elif [ "$plugin" = "bdev" ]; then

					filename=${disks[$n]}

				fi

				sed -i -e "\$afilename=$filename" $BASE_DIR/config.fio

				#Mark numa of n'th disk as "x" to mark it as claimed

				disks_numa[$n]="x"

			fi

			n=$(($n+1))

			# If there is no more disks with numa node same as cpu numa node, switch to other numa node.

			if [ $n -ge $total_disks ]; then

				if [ "$core_numa" = "1" ]; then

					core_numa=0

				else

					core_numa=1

				fi

				n=0

			fi

		done

		echo "" >> $BASE_DIR/config.fio

	done

}



function preconditioning(){

	local dev_name=""

	local filename=""

	local i

	sed -i -e "\$a[preconditioning]" $BASE_DIR/config.fio

	for (( i=0; i < $DISKNO; i++ ))

	do

		dev_name='trtype=PCIe traddr='${disks[i]//:/.}' ns=1'

		filename+=$(printf %s":" "$dev_name")

	done

	echo "** Preconditioning disks, this can take a while, depending on the size of disks."

	run_spdk_nvme_fio "nvme" --filename="$filename" --size=100% --loops=2 --bs=1M\

		--rw=write --iodepth=32

}



function get_results(){

	local reads_pct=$2

	local writes_pct=$((100-$2))



	case "$1" in

		iops)

		iops=$(cat $NVME_FIO_RESULTS | jq -r '.jobs[] | (.read.iops + .write.iops)')

		iops=${iops%.*}

		echo $iops

		;;

		mean_lat_usec)

		mean_lat=$(cat $NVME_FIO_RESULTS | jq -r ".jobs[] | (.read.lat_ns.mean * $reads_pct + .write.lat_ns.mean * $writes_pct)")

		mean_lat=${mean_lat%.*}

		echo $(( $mean_lat/100000 ))

		;;

		p99_lat_usec)

		p99_lat=$(cat $NVME_FIO_RESULTS | jq -r ".jobs[] | (.read.clat_ns.percentile.\"99.000000\" * $reads_pct + .write.clat_ns.percentile.\"99.000000\" * $writes_pct)")

		p99_lat=${p99_lat%.*}

		echo $(( $p99_lat/100000 ))

		;;

		p99_99_lat_usec)

		p99_99_lat=$(cat $NVME_FIO_RESULTS | jq -r ".jobs[] | (.read.clat_ns.percentile.\"99.990000\" * $reads_pct + .write.clat_ns.percentile.\"99.990000\" * $writes_pct)")

		p99_99_lat=${p99_99_lat%.*}

		echo $(( $p99_99_lat/100000 ))

		;;

		stdev_usec)

		stdev=$(cat $NVME_FIO_RESULTS | jq -r ".jobs[] | (.read.clat_ns.stddev * $reads_pct + .write.clat_ns.stddev * $writes_pct)")

		stdev=${stdev%.*}

		echo $(( $stdev/100000 ))

		;;

	esac

}



function run_spdk_nvme_fio(){

	local plugin=$1

	echo "** Running fio test, this can take a while, depending on the run-time and ramp-time setting."

	if [ "$plugin" = "nvme" ]; then

		LD_PRELOAD=$PLUGIN_DIR_NVME/fio_plugin $FIO_BIN $BASE_DIR/config.fio --output-format=json\

		 "${@:2}" --ioengine=spdk

	elif [ "$plugin" = "bdev" ]; then

		LD_PRELOAD=$PLUGIN_DIR_BDEV/fio_plugin $FIO_BIN $BASE_DIR/config.fio --output-format=json\

		 "${@:2}" --ioengine=spdk_bdev --spdk_conf=$BASE_DIR/bdev.conf --spdk_mem=4096

	fi



	sleep 1

}



function usage()

{

	set +x

	[[ ! -z $2 ]] && ( echo "$2"; echo ""; )

	echo "Run NVMe PMD/BDEV performance test. Change options for easier debug and setup configuration"

	echo "Usage: $(basename $1) [options]"

	echo "-h, --help                Print help and exit"

	echo "    --run-time=TIME[s]    Tell fio to run the workload for the specified period of time. [default=$RUNTIME]"

	echo "    --ramp-time=TIME[s]   Fio will run the specified workload for this amount of time before logging any performance numbers. [default=$RAMP_TIME]"

	echo "    --fio-bin=PATH        Path to fio binary. [default=$FIO_BIN]"

	echo "    --fio-plugin=STR      Use bdev or nvme fio_plugin. [default=$PLUGIN]"

	echo "    --max-disk=INT,ALL    Number of disks to test on, this will run multiple workloads with increasing number of disk each run, if =ALL then test on all found disk. [default=$DISKNO]"

	echo "    --disk-no=INT,ALL     Number of disks to test on, this will run one workload on selected number od disks, it discards max-disk setting, if =ALL then test on all found disk"

	echo "    --rwmixread=INT       Percentage of a mixed workload that should be reads. [default=$MIX]"

	echo "    --iodepth=INT         Number of I/Os to keep in flight against the file. [default=$IODEPTH]"

	echo "    --cpu-allowed=INT     Comma-separated list of CPU cores used to run the workload. [default=$CPUS_ALLOWED]"

	echo "    --repeat-no=INT       How many times to repeat each workload. [default=$REPEAT_NO]"

	echo "    --no-preconditioning  Skip preconditioning"

	set -x

}



while getopts 'h-:' optchar; do

	case "$optchar" in

		-)

		case "$OPTARG" in

			help) usage $0; exit 0 ;;

			run-time=*) RUNTIME="${OPTARG#*=}" ;;

			ramp-time=*) RAMP_TIME="${OPTARG#*=}" ;;

			fio-bin=*) FIO_BIN="${OPTARG#*=}" ;;

			max-disk=*) DISKNO="${OPTARG#*=}" ;;

			disk-no=*) DISKNO="${OPTARG#*=}"; ONEWORKLOAD=true ;;

			fio-plugin=*) PLUGIN="${OPTARG#*=}" ;;

			rwmixread=*) MIX="${OPTARG#*=}" ;;

			iodepth=*) IODEPTH="${OPTARG#*=}" ;;

			no-preconditioning) PRECONDITIONING=false ;;

			cpu-allowed=*) CPUS_ALLOWED="${OPTARG#*=}" ;;

			repeat-no=*) REPEAT_NO="${OPTARG#*=}" ;;

			*) usage $0 echo "Invalid argument '$OPTARG'"; exit 1 ;;

		esac

		;;

		h) usage $0; exit 0 ;;

		*) usage $0 "Invalid argument '$optchar'"; exit 1 ;;

	esac

done



trap 'rm -f *.state $BASE_DIR/bdev.conf; print_backtrace' ERR SIGTERM SIGABRT

mkdir -p $BASE_DIR/results

date="$(date +'%m_%d_%Y_%H%M%S')"

disks=($(get_disks nvme))

if [[ $DISKNO == "ALL" ]] || [[ $DISKNO == "all" ]]; then

	DISKNO=${#disks[@]}

elif [[ $DISKNO -gt ${#disks[@]} ]] || [[ ! $DISKNO =~ ^[0-9]+$ ]]; then

	echo "error: Required devices number ($DISKNO) is not a valid number or it's larger than the number of devices found (${#disks[@]})"

	exit 1

fi

test/nvme/perf/config.fio.tmp

0 → 100644
+5 −0
Original line number Diff line number Diff line 
[global]

thread=1

group_reporting=1

direct=1

norandommap=1

test/nvme/perf/run_perf.sh

0 → 100755
+103 −0
Original line number Diff line number Diff line 
#!/usr/bin/env bash



# Automated script that runs NVMe PMD/BDEV performance test.

# This script should be run as root. Please run the scripts/setup.sh before running this script to bind disks to VFIO/UIO driver

# This script takes the following parameters:

# "--run-time" - the run time for the workload in seconds

# "--ramp-time" - Fio will run the specified workload for this amount of time before logging any performance numbers

# "--cpu-allowed" - A comma-separated list of CPU cores used to run the workload. - When the spdk fio plugin is chosen, NVMe devices will

# be aligned to specific core according to their NUMA node. The script will try to align each core with devices matching core's

# on the same NUMA node first but if there are no devices left in the same NUMA node as the CPU Core then it will use devices on the other NUMA node.

# It is important to choose cores that will ensure best NUMA node allocation. For example, on a system with 8 devices on NUMA node

# 0 and 8 devices on NUMA node 1, cores 0-27 on numa node 0 and 28-55 on numa node 1, if test uses 16 disk and four cores

# then "--cpu-allowed=1,2,28,29" results in a NUMA-balanced configuration with 4 devices on each CPU core.

# However, if the test will use 10 CPU cores, then best option would be "--cpu-allowed=1,2,3,4,28,29,30,31,32,33" because cores 1-4 will be aligned with

# 2 devices on numa0 per core, cores 28-29 will be aligned with 2 devices on numa1 per core and cores 30-33 with 1 device on numa1 per core.

# "--iodepth" - Number of I/Os to keep in flight per devices.

# "--fio-plugin" - The SPDK fio plugin used in this test - bdev or nvme

# "--no-preconditioning" - skip preconditioning - Normally the script will precondition disks to put them in a steady state.

# However, preconditioning could be skipped, for example preconditiong has been already made and workload was 100% reads.

# "--disk-no" - use specified number of disks for test.

# "--repeat-no" Repeat each workolad specified number of times.

# An Example Performance Test Run

# "./spdk/test/perf/run_perf.sh --run-time=600 --ramp-time=60 --cpu-allowed=28 --fio-bin=/usr/src/fio/fio\

#  --rwmixread=100 --iodepth=256 --fio-plugin=bdev --no-preconditioning --disk-no=6"

# This command will run test using fio plugin for 600 seconds, 60 sec of ram time, randrw job with

# 100% reads with io depth 256 per disk, on 6 devices and skips preconditioning. Cpu core used for this test is

# core no 28.

BASE_DIR=$(readlink -f $(dirname $0))

. $BASE_DIR/common.sh

if [ $PLUGIN = "bdev" ]; then

	$ROOT_DIR/scripts/gen_nvme.sh >> $BASE_DIR/bdev.conf

fi



disk_names=$(get_disks $PLUGIN)

disks_numa=$(get_numa_node $PLUGIN "$disk_names")

cores=$(get_cores "$CPUS_ALLOWED")

no_cores=($cores)

no_cores=${#no_cores[@]}



if $PRECONDITIONING; then

	cp $BASE_DIR/config.fio.tmp $BASE_DIR/config.fio

	preconditioning

	rm -f $BASE_DIR/config.fio

fi



result_dir=perf_results_${BLK_SIZE}BS_${IODEPTH}QD_${RW}_${MIX}MIX_${PLUGIN}_${date}

mkdir $BASE_DIR/results/$result_dir

result_file=$BASE_DIR/results/$result_dir/perf_results_${BLK_SIZE}BS_${IODEPTH}QD_${RW}_${MIX}MIX_${PLUGIN}_${date}.csv

unset iops_disks mean_lat_disks p99_lat_disks p99_99_lat_disks stdev_disks

echo "run-time,ramp-time,fio-plugin,QD,block-size,num-cpu-cores,workload,workload-mix" > $result_file

printf "%s,%s,%s,%s,%s,%s,%s,%s\n" $RUNTIME $RAMP_TIME $PLUGIN $IODEPTH $BLK_SIZE $no_cores $RW $MIX >> $result_file

echo "num_of_disks,iops,avg_lat[usec],p99[usec],p99.99[usec],stdev[usec]" >> $result_file

#Run each workolad $REPEAT_NO times

for (( j=0; j < $REPEAT_NO; j++ ))

do

	#Start with $DISKNO disks and remove 2 disks for each run to avoid preconditioning before each run.

	for (( k=$DISKNO; k >= 1; k-=2 ))

	do

		cp $BASE_DIR/config.fio.tmp $BASE_DIR/config.fio

		echo "" >> $BASE_DIR/config.fio

		#The SPDK fio plugin supports submitting/completing I/Os to multiple SSDs from a single thread.

		#Therefore, the per thread queue depth is set to the desired IODEPTH/device X the number of devices per thread.

		qd=$(( $IODEPTH * $k ))

		filename=$(create_fio_config $k $PLUGIN "$disk_names" "$disks_numa" "$cores")

		desc="Running Test: Blocksize=${BLK_SIZE} Workload=$RW MIX=${MIX} qd=${IODEPTH} fio_plugin=$PLUGIN"

		run_spdk_nvme_fio $PLUGIN "--runtime=$RUNTIME" "--ramp_time=$RAMP_TIME" "--bs=$BLK_SIZE"\

		 "--rw=$RW" "--rwmixread=$MIX" "--iodepth=$qd" "--output=$NVME_FIO_RESULTS" "--time_based=1"\

		 "--description=$desc"



		#Store values for every number of used disks

		iops_disks[$k]=$((${iops_disks[$k]} + $(get_results iops $MIX)))

		mean_lat_disks_usec[$k]=$((${mean_lat_disks_usec[$k]} + $(get_results mean_lat_usec $MIX)))

		p99_lat_disks_usec[$k]=$((${p99_lat_disks_usec[$k]} + $(get_results p99_lat_usec $MIX)))

		p99_99_lat_disks_usec[$k]=$((${p99_99_lat_disks_usec[$k]} + $(get_results p99_99_lat_usec $MIX)))

		stdev_disks_usec[$k]=$((${stdev_disks_usec[$k]} + $(get_results stdev_usec $MIX)))

		cp $NVME_FIO_RESULTS $BASE_DIR/results/$result_dir/perf_results_${MIX}_${PLUGIN}_${no_cores}cpus_${date}_${k}_disks_${j}.json

		cp $BASE_DIR/config.fio $BASE_DIR/results/$result_dir/config_${MIX}_${PLUGIN}_${no_cores}cpus_${date}_${k}_disks_${j}.fio

		rm -f $BASE_DIR/config.fio



		#if tested on only one number of disk

		if $ONEWORKLOAD; then

			break

		fi

	done

done

#Write results to csv file

for (( k=$DISKNO; k >= 1; k-=2 ))

do

	iops_disks[$k]=$((${iops_disks[$k]} / 3))

	mean_lat_disks_usec[$k]=$((${mean_lat_disks_usec[$k]} / 3))

	p99_lat_disks_usec[$k]=$((${p99_lat_disks_usec[$k]} / 3))

	p99_99_lat_disks_usec[$k]=$((${p99_99_lat_disks_usec[$k]} / 3))

	stdev_disks_usec[$k]=$((${stdev_disks_usec[$k]} / 3))



	printf "%s,%s,%s,%s,%s,%s\n" ${k} ${iops_disks[$k]} ${mean_lat_disks_usec[$k]} ${p99_lat_disks_usec[$k]}\

	 ${p99_99_lat_disks_usec[$k]} ${stdev_disks_usec[$k]} >> $result_file



	#if tested on only one numeber of disk

	if $ONEWORKLOAD; then

		break

	fi

done

rm -f $BASE_DIR/bdev.conf $BASE_DIR/config.fio