Merge pull request #360 from gwsystems/359-perf-window

refactor: perf_window
master
Sean McBride 2 years ago committed by GitHub
commit e011912bdf
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23

@ -21,8 +21,8 @@ perf_window_initialize(struct perf_window *perf_window)
lock_init(&perf_window->lock);
perf_window->count = 0;
memset(perf_window->by_duration, 0, sizeof(struct execution_node) * PERF_WINDOW_BUFFER_SIZE);
memset(perf_window->by_termination, 0, sizeof(uint16_t) * PERF_WINDOW_BUFFER_SIZE);
memset(perf_window->by_duration, 0, sizeof(struct execution_node) * perf_window_capacity);
memset(perf_window->by_termination, 0, sizeof(uint16_t) * perf_window_capacity);
}
@ -38,8 +38,8 @@ perf_window_swap(struct perf_window *perf_window, uint16_t first_by_duration_idx
{
assert(lock_is_locked(&perf_window->lock));
assert(perf_window != NULL);
assert(first_by_duration_idx >= 0 && first_by_duration_idx < PERF_WINDOW_BUFFER_SIZE);
assert(second_by_duration_idx >= 0 && second_by_duration_idx < PERF_WINDOW_BUFFER_SIZE);
assert(first_by_duration_idx < perf_window_capacity);
assert(second_by_duration_idx < perf_window_capacity);
uint16_t first_by_termination_idx = perf_window->by_duration[first_by_duration_idx].by_termination_idx;
uint16_t second_by_termination_idx = perf_window->by_duration[second_by_duration_idx].by_termination_idx;
@ -51,11 +51,11 @@ perf_window_swap(struct perf_window *perf_window, uint16_t first_by_duration_idx
uint64_t first_execution_time = perf_window->by_duration[first_by_duration_idx].execution_time;
uint64_t second_execution_time = perf_window->by_duration[second_by_duration_idx].execution_time;
/* Swap Indices in Buffer*/
/* Swap indices */
perf_window->by_termination[first_by_termination_idx] = second_by_duration_idx;
perf_window->by_termination[second_by_termination_idx] = first_by_duration_idx;
/* Swap by_termination_idx */
/* Swap nodes */
struct execution_node tmp_node = perf_window->by_duration[first_by_duration_idx];
perf_window->by_duration[first_by_duration_idx] = perf_window->by_duration[second_by_duration_idx];
perf_window->by_duration[second_by_duration_idx] = tmp_node;
@ -67,66 +67,73 @@ perf_window_swap(struct perf_window *perf_window, uint16_t first_by_duration_idx
== second_execution_time);
}
static inline void
perf_window_fill(struct perf_window *perf_window, uint64_t newest_execution_time)
{
for (uint16_t i = 0; i < perf_window_capacity; i++) {
perf_window->by_termination[i] = i;
perf_window->by_duration[i] = (struct execution_node){ .execution_time = newest_execution_time,
.by_termination_idx = i };
}
perf_window->count = perf_window_capacity;
}
/**
* Adds a new value to the perf window
* Adds newest_execution_time to the perf window
* Not intended to be called directly!
* @param perf_window
* @param value
* @param newest_execution_time
*/
static inline void
perf_window_add(struct perf_window *perf_window, uint64_t value)
perf_window_add(struct perf_window *perf_window, uint64_t newest_execution_time)
{
assert(perf_window != NULL);
/* Assumption: A successful invocation should run for a non-zero amount of time */
assert(newest_execution_time > 0);
uint16_t idx_of_oldest;
uint16_t idx_to_replace;
uint64_t previous_execution_time;
bool check_up;
if (unlikely(!lock_is_locked(&perf_window->lock))) panic("lock not held when calling perf_window_add\n");
/* A successful invocation should run for a non-zero amount of time */
assert(value > 0);
/* If count is 0, then fill entire array with initial execution times */
/* If perf window is empty, fill all elements with newest_execution_time */
if (perf_window->count == 0) {
for (int i = 0; i < PERF_WINDOW_BUFFER_SIZE; i++) {
perf_window->by_termination[i] = i;
perf_window->by_duration[i] = (struct execution_node){ .execution_time = value,
.by_termination_idx = i };
}
perf_window->count = PERF_WINDOW_BUFFER_SIZE;
perf_window_fill(perf_window, newest_execution_time);
goto done;
}
/* Otherwise, replace the oldest value, and then sort */
idx_of_oldest = perf_window->by_termination[perf_window->count % PERF_WINDOW_BUFFER_SIZE];
check_up = value > perf_window->by_duration[idx_of_oldest].execution_time;
perf_window->by_duration[idx_of_oldest].execution_time = value;
/* If full, replace the oldest execution_time. Save the old execution time to know which direction to swap */
idx_to_replace = perf_window->by_termination[perf_window->count % perf_window_capacity];
previous_execution_time = perf_window->by_duration[idx_to_replace].execution_time;
perf_window->by_duration[idx_to_replace].execution_time = newest_execution_time;
if (check_up) {
for (uint16_t i = idx_of_oldest;
i + 1 < PERF_WINDOW_BUFFER_SIZE
/* At this point, the by_duration array is partially sorted. The node we overwrote needs to be shifted left or
* right. We can determine which direction to shift by comparing with the previous execution time. */
if (newest_execution_time > previous_execution_time) {
for (uint16_t i = idx_to_replace;
i + 1 < perf_window_capacity
&& perf_window->by_duration[i + 1].execution_time < perf_window->by_duration[i].execution_time;
i++) {
perf_window_swap(perf_window, i, i + 1);
}
} else {
for (int i = idx_of_oldest;
i - 1 >= 0
for (uint16_t i = idx_to_replace;
i >= 1
&& perf_window->by_duration[i - 1].execution_time > perf_window->by_duration[i].execution_time;
i--) {
perf_window_swap(perf_window, i, i - 1);
}
}
/* The idx that we replaces should still point to the same value */
assert(perf_window->by_duration[perf_window->by_termination[perf_window->count % PERF_WINDOW_BUFFER_SIZE]]
/* The idx that we replaces should still point to the same newest_execution_time */
assert(perf_window->by_duration[perf_window->by_termination[perf_window->count % perf_window_capacity]]
.execution_time
== value);
== newest_execution_time);
/* The by_duration array should be ordered by execution time */
#ifndef NDEBUG
for (int i = 1; i < PERF_WINDOW_BUFFER_SIZE; i++) {
for (int i = 1; i < perf_window_capacity; i++) {
assert(perf_window->by_duration[i - 1].execution_time <= perf_window->by_duration[i].execution_time);
}
#endif
@ -151,10 +158,10 @@ perf_window_get_percentile(struct perf_window *perf_window, uint8_t percentile,
assert(percentile >= 50 && percentile <= 99);
assert(perf_window->count > 0);
if (likely(perf_window->count >= PERF_WINDOW_BUFFER_SIZE))
return perf_window->by_duration[precomputed_index].execution_time;
int idx = precomputed_index;
if (unlikely(perf_window->count < perf_window_capacity)) idx = perf_window->count * percentile / 100;
return perf_window->by_duration[perf_window->count * percentile / 100].execution_time;
return perf_window->by_duration[idx].execution_time;
}
/**

@ -1,15 +1,19 @@
#pragma once
#include <assert.h>
#include <stdint.h>
#include "lock.h"
/* Should be Power of 2! */
#define PERF_WINDOW_BUFFER_SIZE 16
enum
{
perf_window_capacity = 32
};
static_assert(perf_window_capacity && !(perf_window_capacity & (perf_window_capacity - 1)),
"perf_window_capacity must be power of 2!");
#if ((PERF_WINDOW_BUFFER_SIZE == 0) || (PERF_WINDOW_BUFFER_SIZE & (PERF_WINDOW_BUFFER_SIZE - 1)) != 0)
#error "PERF_WINDOW_BUFFER_SIZE must be power of 2!"
#endif
static_assert(perf_window_capacity <= UINT16_MAX, "perf_window_capacity must be indexable by a 16-bit unsigned int");
/*
* The by_duration array sorts the last N executions by execution time
@ -24,8 +28,8 @@ struct execution_node {
};
struct perf_window {
struct execution_node by_duration[PERF_WINDOW_BUFFER_SIZE];
uint16_t by_termination[PERF_WINDOW_BUFFER_SIZE];
struct execution_node by_duration[perf_window_capacity];
uint16_t by_termination[perf_window_capacity];
uint64_t count;
lock_t lock;
};

@ -4,6 +4,7 @@
#include <stdio.h>
#include <stdlib.h>
#include "admissions_control.h"
#include "runtime.h"
#include "scheduler_options.h"
@ -100,12 +101,12 @@ route_config_validate(struct route_config *config, bool *did_set)
if (config->admissions_percentile > 99 || config->admissions_percentile < 50) {
fprintf(stderr, "admissions-percentile must be > 50 and <= 99 but was %u\n",
route_config->admissions_percentile);
config->admissions_percentile);
return -1;
}
/* If the ratio is too big, admissions control is too coarse */
uint32_t ratio = route_config->relative_deadline_us / route_config->expected_execution_us;
uint32_t ratio = config->relative_deadline_us / config->expected_execution_us;
if (ratio > ADMISSIONS_CONTROL_GRANULARITY) {
fprintf(stderr,
"Ratio of Deadline to Execution time cannot exceed admissions control "

@ -3,6 +3,9 @@
#include "admissions_control.h"
#include "debuglog.h"
#include "likely.h"
#include "panic.h"
#include "runtime.h"
/*
* Unitless estimate of the instantaneous fraction of system capacity required to complete all previously

@ -1,5 +1,7 @@
#include "admissions_control.h"
#include "admissions_info.h"
#include "debuglog.h"
#include "perf_window.h"
/**
* Initializes perf window
@ -22,7 +24,7 @@ admissions_info_initialize(struct admissions_info *admissions_info, uint8_t perc
if (unlikely(percentile < 50 || percentile > 99)) panic("Invalid admissions percentile");
admissions_info->percentile = percentile;
admissions_info->control_index = PERF_WINDOW_BUFFER_SIZE * percentile / 100;
admissions_info->control_index = perf_window_capacity * percentile / 100;
#ifdef LOG_ADMISSIONS_CONTROL
debuglog("Percentile: %u\n", admissions_info->percentile);
debuglog("Control Index: %d\n", admissions_info->control_index);

@ -38,16 +38,16 @@ run_samples() {
# 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" ../../../runtime/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"
local -i perf_window_capacity
if ! perf_window_capacity=$(grep "perf_window_capacity =" < "$perf_window_path" | cut -d\ -f3); then
printf "Failed to scrape perf_window_capacity from ../../include/perf_window.h\n"
printf "Defaulting to 16\n"
perf_window_buffer_size=16
perf_window_capacity=16
fi
local -ir perf_window_buffer_size
local -ir perf_window_capacity
printf "Running Samples: "
hey -disable-compression -disable-keepalive -disable-redirects -n "$perf_window_buffer_size" -c "$perf_window_buffer_size" -q 200 -cpus 3 -o csv -m GET "http://${hostname}:10000" 1> /dev/null 2> /dev/null || {
hey -disable-compression -disable-keepalive -disable-redirects -n "$perf_window_capacity" -c "$perf_window_capacity" -q 200 -cpus 3 -o csv -m GET "http://${hostname}:10000/empty" 1> /dev/null 2> /dev/null || {
printf "[ERR]\n"
panic "samples failed"
return 1
@ -79,7 +79,7 @@ run_experiments() {
printf "Running Experiments:\n"
for conn in ${concurrency[*]}; do
printf "\t%d Concurrency: " "$conn"
hey -disable-compression -disable-keepalive -disable-redirects -n "$iterations" -c "$conn" -cpus 2 -o csv -m GET "http://$hostname:10000" > "$results_directory/con$conn.csv" 2> /dev/null || {
hey -disable-compression -disable-keepalive -disable-redirects -n "$iterations" -c "$conn" -cpus 2 -o csv -m GET "http://$hostname:10000/empty" > "$results_directory/con$conn.csv" 2> /dev/null || {
printf "[ERR]\n"
panic "experiment failed"
return 1

@ -41,22 +41,22 @@ run_samples() {
# 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" ../../../runtime/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 ../../../runtime/include/perf_window.h\n"
local -i perf_window_capacity
if ! perf_window_capacity=$(grep "perf_window_capacity =" < "$perf_window_path" | cut -d\ -f3); then
printf "Failed to scrape perf_window_capacity from ../../../runtime/include/perf_window.h\n"
printf "Defaulting to 16\n"
perf_window_buffer_size=16
perf_window_capacity=16
fi
local -ir perf_window_buffer_size
local -ir perf_window_capacity
printf "Running Samples: "
hey -disable-compression -disable-keepalive -disable-redirects -n "$perf_window_buffer_size" -c "$perf_window_buffer_size" -cpus 3 -t 0 -o csv -m GET -d "40\n" "http://${hostname}:10010/fib2" 1> /dev/null 2> /dev/null || {
hey -disable-compression -disable-keepalive -disable-redirects -n "$perf_window_capacity" -c "$perf_window_capacity" -cpus 3 -t 0 -o csv -m GET -d "40\n" "http://${hostname}:10010/fib2" 1> /dev/null 2> /dev/null || {
printf "[ERR]\n"
panic "fib40 samples failed with $?"
return 1
}
hey -disable-compression -disable-keepalive -disable-redirects -n "$perf_window_buffer_size" -c "$perf_window_buffer_size" -cpus 3 -t 0 -o csv -m GET -d "10\n" "http://${hostname}:100010/fib" 1> /dev/null 2> /dev/null || {
hey -disable-compression -disable-keepalive -disable-redirects -n "$perf_window_capacity" -c "$perf_window_capacity" -cpus 3 -t 0 -o csv -m GET -d "10\n" "http://${hostname}:100010/fib" 1> /dev/null 2> /dev/null || {
printf "[ERR]\n"
panic "fib10 samples failed with $?"
return 1

@ -38,22 +38,22 @@ run_samples() {
# 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" ../../runtime/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 ../../runtime/include/perf_window.h\n"
local -i perf_window_capacity
if ! perf_window_capacity=$(grep "perf_window_capacity =" < "$perf_window_path" | cut -d\ -f3); then
printf "Failed to scrape perf_window_capacity from ../../runtime/include/perf_window.h\n"
printf "Defaulting to 16\n"
perf_window_buffer_size=16
perf_window_capacity=16
fi
local -ir perf_window_buffer_size
local -ir perf_window_capacity
printf "Running Samples: "
hey -disable-compression -disable-keepalive -disable-redirects -n "$perf_window_buffer_size" -c "$perf_window_buffer_size" -cpus 3 -t 0 -o csv -m GET -d "40\n" "http://${hostname}:10000/fibonacci_10" 1> /dev/null 2> /dev/null || {
hey -disable-compression -disable-keepalive -disable-redirects -n "$perf_window_capacity" -c "$perf_window_capacity" -cpus 3 -t 0 -o csv -m GET -d "40\n" "http://${hostname}:10000/fibonacci_10" 1> /dev/null 2> /dev/null || {
printf "[ERR]\n"
panic "fibonacci_40 samples failed with $?"
return 1
}
hey -disable-compression -disable-keepalive -disable-redirects -n "$perf_window_buffer_size" -c "$perf_window_buffer_size" -cpus 3 -t 0 -o csv -m GET -d "10\n" "http://${hostname}:10000/fibonacci_10" 1> /dev/null 2> /dev/null || {
hey -disable-compression -disable-keepalive -disable-redirects -n "$perf_window_capacity" -c "$perf_window_capacity" -cpus 3 -t 0 -o csv -m GET -d "10\n" "http://${hostname}:10000/fibonacci_10" 1> /dev/null 2> /dev/null || {
printf "[ERR]\n"
panic "fibonacci_10 samples failed with $?"
return 1

Loading…
Cancel
Save