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.
sledge/runtime/include/perf_window.h

186 lines
6.1 KiB

#pragma once
#include <stdint.h>
#include "lock.h"
#include "runtime.h"
#include "worker_thread.h"
/* Should be Power of 2! */
#define PERF_WINDOW_BUFFER_SIZE 16
#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
/*
* The by_duration array sorts the last N executions by execution time
* The by_termination array acts as a circular buffer that maps to indices in the by_duration array
*
* by_termination ensures that the when the circular buffer is full, the oldest data in both arrays is
* overwritten, providing a sorted circular buffer
*/
struct execution_node {
uint64_t execution_time;
uint16_t by_termination_idx; /* Reverse idx of the associated by_termination bin. Used for swaps! */
};
struct perf_window {
struct execution_node by_duration[PERF_WINDOW_BUFFER_SIZE];
uint16_t by_termination[PERF_WINDOW_BUFFER_SIZE];
uint64_t count;
lock_t lock;
};
/**
* Initializes perf window
* @param self
*/
static inline void
perf_window_initialize(struct perf_window *self)
{
assert(self != NULL);
LOCK_INIT(&self->lock);
self->count = 0;
memset(&self->by_duration, 0, sizeof(struct execution_node) * PERF_WINDOW_BUFFER_SIZE);
memset(&self->by_termination, 0, sizeof(uint16_t) * PERF_WINDOW_BUFFER_SIZE);
}
/**
* Swaps two execution nodes in the by_duration array, including updating the indices in the by_termination circular
* buffer
* @param self
* @param first_by_duration_idx
* @param second_by_duration_idx
*/
static inline void
perf_window_swap(struct perf_window *self, uint16_t first_by_duration_idx, uint16_t second_by_duration_idx)
{
assert(LOCK_IS_LOCKED(&self->lock));
assert(self != 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);
uint16_t first_by_termination_idx = self->by_duration[first_by_duration_idx].by_termination_idx;
uint16_t second_by_termination_idx = self->by_duration[second_by_duration_idx].by_termination_idx;
/* The execution node's by_termination_idx points to a by_termination cell equal to its own by_duration index */
assert(self->by_termination[first_by_termination_idx] == first_by_duration_idx);
assert(self->by_termination[second_by_termination_idx] == second_by_duration_idx);
uint64_t first_execution_time = self->by_duration[first_by_duration_idx].execution_time;
uint64_t second_execution_time = self->by_duration[second_by_duration_idx].execution_time;
/* Swap Indices in Buffer*/
self->by_termination[first_by_termination_idx] = second_by_duration_idx;
self->by_termination[second_by_termination_idx] = first_by_duration_idx;
/* Swap by_termination_idx */
struct execution_node tmp_node = self->by_duration[first_by_duration_idx];
self->by_duration[first_by_duration_idx] = self->by_duration[second_by_duration_idx];
self->by_duration[second_by_duration_idx] = tmp_node;
/* The circular by_termination indices should always point to the same execution times across all swaps */
assert(self->by_duration[self->by_termination[first_by_termination_idx]].execution_time
== first_execution_time);
assert(self->by_duration[self->by_termination[second_by_termination_idx]].execution_time
== second_execution_time);
}
/**
* Adds a new value to the perf window
* Not intended to be called directly!
* @param self
* @param value
*/
static inline void
perf_window_add(struct perf_window *self, uint64_t value)
{
assert(self != NULL);
if (unlikely(!LOCK_IS_LOCKED(&self->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 (self->count == 0) {
for (int i = 0; i < PERF_WINDOW_BUFFER_SIZE; i++) {
self->by_termination[i] = i;
self->by_duration[i] = (struct execution_node){ .execution_time = value,
.by_termination_idx = i };
}
self->count = PERF_WINDOW_BUFFER_SIZE;
goto done;
}
/* Otherwise, replace the oldest value, and then sort */
uint16_t idx_of_oldest = self->by_termination[self->count % PERF_WINDOW_BUFFER_SIZE];
bool check_up = value > self->by_duration[idx_of_oldest].execution_time;
self->by_duration[idx_of_oldest].execution_time = value;
if (check_up) {
for (uint16_t i = idx_of_oldest;
i + 1 < PERF_WINDOW_BUFFER_SIZE
&& self->by_duration[i + 1].execution_time < self->by_duration[i].execution_time;
i++) {
perf_window_swap(self, i, i + 1);
}
} else {
for (int i = idx_of_oldest;
i - 1 >= 0 && self->by_duration[i - 1].execution_time > self->by_duration[i].execution_time; i--) {
perf_window_swap(self, i, i - 1);
}
}
/* The idx that we replaces should still point to the same value */
assert(self->by_duration[self->by_termination[self->count % PERF_WINDOW_BUFFER_SIZE]].execution_time == value);
/* The by_duration array should be ordered by execution time */
#ifndef NDEBUG
for (int i = 1; i < PERF_WINDOW_BUFFER_SIZE; i++) {
assert(self->by_duration[i - 1].execution_time <= self->by_duration[i].execution_time);
}
#endif
self->count++;
done:
return;
}
/**
* Returns pXX execution time
* @param self
* @param percentile represented by int between 50 and 99
* @param precomputed_index memoized index for quick lookup when by_duration is full
* @returns execution time
*/
static inline uint64_t
perf_window_get_percentile(struct perf_window *self, int percentile, int precomputed_index)
{
assert(self != NULL);
assert(percentile >= 50 && percentile <= 99);
int size = self->count;
assert(size > 0);
if (likely(size >= PERF_WINDOW_BUFFER_SIZE)) return self->by_duration[precomputed_index].execution_time;
return self->by_duration[size * percentile / 100].execution_time;
}
/**
* Returns the total count of executions
* @returns total count
*/
static inline uint64_t
perf_window_get_count(struct perf_window *self)
{
assert(self != NULL);
return self->count;
}