You can not select more than 25 topics Topics must start with a letter or number, can include dashes ('-') and can be up to 35 characters long.

265 lines
8.4 KiB

#!/bin/bash
# This experiment is intended to document how the level of concurrent requests influence the latency, throughput, and success/failure rate
# Success - The percentage of requests that complete by their deadlines
# TODO: Does this handle non-200s?
# Throughput - The mean number of successful requests per second
# Latency - the rount-trip resonse time (unit?) of successful requests at the p50, p90, p99, and p100 percetiles
# Add bash_libraries directory to path
__run_sh__base_path="$(dirname "$(realpath --logical "${BASH_SOURCE[0]}")")"
__run_sh__bash_libraries_relative_path="../bash_libraries"
__run_sh__bash_libraries_absolute_path=$(cd "$__run_sh__base_path" && cd "$__run_sh__bash_libraries_relative_path" && pwd)
export PATH="$__run_sh__bash_libraries_absolute_path:$PATH"
source csv_to_dat.sh || exit 1
source framework.sh || exit 1
# source generate_gnuplots.sh || exit 1
source get_result_count.sh || exit 1
source panic.sh || exit 1
source path_join.sh || exit 1
if ! command -v hey > /dev/null; then
echo "hey is not present."
exit 1
fi
# Sends requests until the per-module perf window buffers are full
# This ensures that Sledge has accurate estimates of execution time
run_samples() {
if (($# != 1)); then
panic "invalid number of arguments \"$1\""
return 1
elif [[ -z "$1" ]]; then
panic "hostname \"$1\" was empty"
return 1
fi
local hostname="${1}"
# Scrape the perf window size from the source if possible
# TODO: Make a util function
local -r perf_window_path="$(path_join "$__run_sh__base_path" ../../include/perf_window_t.h)"
local -i perf_window_buffer_size
if ! perf_window_buffer_size=$(grep "#define PERF_WINDOW_BUFFER_SIZE" < "$perf_window_path" | cut -d\ -f3); then
printf "Failed to scrape PERF_WINDOW_BUFFER_SIZE from ../../include/perf_window.h\n"
printf "Defaulting to 16\n"
perf_window_buffer_size=16
fi
local -ir perf_window_buffer_size
printf "Running Samples: "
hey -n "$perf_window_buffer_size" -c "$perf_window_buffer_size" -cpus 3 -t 0 -o csv -m GET -d "40\n" "http://${hostname}:10040" 1> /dev/null 2> /dev/null || {
printf "[ERR]\n"
panic "fib40 samples failed with $?"
return 1
}
hey -n "$perf_window_buffer_size" -c "$perf_window_buffer_size" -cpus 3 -t 0 -o csv -m GET -d "10\n" "http://${hostname}:100010" 1> /dev/null 2> /dev/null || {
printf "[ERR]\n"
panic "fib10 samples failed with $?"
return 1
}
printf "[OK]\n"
return 0
}
# Execute the fib10 and fib40 experiments sequentially and concurrently
# $1 (hostname)
# $2 (results_directory) - a directory where we will store our results
run_experiments() {
if (($# != 2)); then
panic "invalid number of arguments \"$1\""
return 1
elif [[ -z "$1" ]]; then
panic "hostname \"$1\" was empty"
return 1
elif [[ ! -d "$2" ]]; then
panic "directory \"$2\" does not exist"
return 1
fi
local hostname="$1"
local results_directory="$2"
# The duration in seconds that we want the client to send requests
local -ir duration_sec=15
# The duration in seconds that the low priority task should run before the high priority task starts
local -ir offset=5
printf "Running Experiments\n"
# Run each separately
printf "\tfib40: "
hey -z ${duration_sec}s -cpus 4 -c 100 -t 0 -o csv -m GET -d "40\n" "http://$hostname:10040" > "$results_directory/fib40.csv" 2> /dev/null || {
printf "[ERR]\n"
panic "fib40 failed"
return 1
}
get_result_count "$results_directory/fib40.csv" || {
printf "[ERR]\n"
panic "fib40 unexpectedly has zero requests"
return 1
}
printf "[OK]\n"
printf "\tfib10: "
hey -z ${duration_sec}s -cpus 4 -c 100 -t 0 -o csv -m GET -d "10\n" "http://$hostname:10010" > "$results_directory/fib10.csv" 2> /dev/null || {
printf "[ERR]\n"
panic "fib10 failed"
return 1
}
get_result_count "$results_directory/fib10.csv" || {
printf "[ERR]\n"
panic "fib10 unexpectedly has zero requests"
return 1
}
printf "[OK]\n"
# Run concurrently
# The lower priority has offsets to ensure it runs the entire time the high priority is trying to run
# This asynchronously trigger jobs and then wait on their pids
local fib40_con_PID
local fib10_con_PID
hey -z $((duration_sec + 2 * offset))s -cpus 2 -c 100 -t 0 -o csv -m GET -d "40\n" "http://${hostname}:10040" > "$results_directory/fib40_con.csv" 2> /dev/null &
fib40_con_PID="$!"
sleep $offset
hey -z "${duration_sec}s" -cpus 2 -c 100 -t 0 -o csv -m GET -d "10\n" "http://${hostname}:10010" > "$results_directory/fib10_con.csv" 2> /dev/null &
fib10_con_PID="$!"
wait -f "$fib10_con_PID" || {
printf "\tfib10_con: [ERR]\n"
panic "failed to wait -f ${fib10_con_PID}"
return 1
}
get_result_count "$results_directory/fib10_con.csv" || {
printf "\tfib10_con: [ERR]\n"
panic "fib10_con has zero requests. This might be because fib40_con saturated the runtime"
return 1
}
printf "\tfib10_con: [OK]\n"
wait -f "$fib40_con_PID" || {
printf "\tfib40_con: [ERR]\n"
panic "failed to wait -f ${fib40_con_PID}"
return 1
}
get_result_count "$results_directory/fib40_con.csv" || {
printf "\tfib40_con: [ERR]\n"
panic "fib40_con has zero requests."
return 1
}
printf "\tfib40_con: [OK]\n"
return 0
}
# Process the experimental results and generate human-friendly results for success rate, throughput, and latency
process_results() {
if (($# != 1)); then
error_msg "invalid number of arguments ($#, expected 1)"
return 1
elif ! [[ -d "$1" ]]; then
error_msg "directory $1 does not exist"
return 1
fi
local -r results_directory="$1"
printf "Processing Results: "
# Write headers to CSVs
printf "Payload,Success_Rate\n" >> "$results_directory/success.csv"
printf "Payload,Throughput\n" >> "$results_directory/throughput.csv"
printf "Payload,p50,p90,p99,p100\n" >> "$results_directory/latency.csv"
# The four types of results that we are capturing.
# fib10 and fib 40 are run sequentially.
# fib10_con and fib40_con are run concurrently
local -ar payloads=(fib10 fib10_con fib40 fib40_con)
# The deadlines for each of the workloads
# TODO: Scrape these from spec.json
local -Ar deadlines_ms=(
[fib10]=2
[fib40]=3000
)
for payload in "${payloads[@]}"; do
# Strip the _con suffix when getting the deadline
local -i deadline=${deadlines_ms[${payload/_con/}]}
# Calculate Success Rate for csv (percent of requests that return 200 within deadline)
awk -F, '
$7 == 200 && ($1 * 1000) <= '"$deadline"' {ok++}
END{printf "'"$payload"',%3.5f\n", (ok / (NR - 1) * 100)}
' < "$results_directory/$payload.csv" >> "$results_directory/success.csv"
# Filter on 200s, convert from s to ms, and sort
awk -F, '$7 == 200 {print ($1 * 1000)}' < "$results_directory/$payload.csv" \
| sort -g > "$results_directory/$payload-response.csv"
# Get Number of 200s
oks=$(wc -l < "$results_directory/$payload-response.csv")
((oks == 0)) && continue # If all errors, skip line
# We determine duration by looking at the timestamp of the last complete request
# TODO: Should this instead just use the client-side synthetic duration_sec value?
duration=$(tail -n1 "$results_directory/$payload.csv" | cut -d, -f8)
# Throughput is calculated as the mean number of successful requests per second
throughput=$(echo "$oks/$duration" | bc)
printf "%s,%f\n" "$payload" "$throughput" >> "$results_directory/throughput.csv"
# Generate Latency Data for csv
awk '
BEGIN {
sum = 0
p50 = int('"$oks"' * 0.5)
p90 = int('"$oks"' * 0.9)
p99 = int('"$oks"' * 0.99)
p100 = '"$oks"'
printf "'"$payload"',"
}
NR==p50 {printf "%1.4f,", $0}
NR==p90 {printf "%1.4f,", $0}
NR==p99 {printf "%1.4f,", $0}
NR==p100 {printf "%1.4f\n", $0}
' < "$results_directory/$payload-response.csv" >> "$results_directory/latency.csv"
# Delete scratch file used for sorting/counting
# rm -rf "$results_directory/$payload-response.csv"
done
# Transform csvs to dat files for gnuplot
csv_to_dat "$results_directory/success.csv" "$results_directory/throughput.csv" "$results_directory/latency.csv"
# Generate gnuplots. Commented out because we don't have *.gnuplots defined
# generate_gnuplots "$results_directory" "$__run_sh__base_path" || {
# printf "[ERR]\n"
# panic "failed to generate gnuplots"
# }
printf "[OK]\n"
return 0
}
# Expected Symbol used by the framework
experiment_main() {
local -r target_hostname="$1"
local -r results_directory="$2"
run_samples "$target_hostname" || return 1
run_experiments "$target_hostname" "$results_directory" || return 1
process_results "$results_directory" || return 1
return 0
}
main "$@"