dendrobates-t-azureus/cache_utils/src/calibration.rs

445 lines
12 KiB
Rust
Raw Normal View History

#![allow(clippy::missing_safety_doc)]
use crate::{flush, maccess, rdtsc_fence};
use core::arch::x86_64 as arch_x86;
#[cfg(feature = "no_std")]
2020-04-01 16:31:06 +02:00
use polling_serial::{serial_print as print, serial_println as println};
2020-02-17 15:28:10 +01:00
#[derive(Ord, PartialOrd, Eq, PartialEq)]
pub enum Verbosity {
NoOutput,
Thresholds,
RawResult,
Debug,
}
pub struct HistParams {
iterations: u32,
bucket_size: usize,
bucket_number: usize,
}
2020-02-17 15:28:10 +01:00
extern crate alloc;
2020-04-01 16:31:06 +02:00
use crate::calibration::Verbosity::*;
use crate::complex_addressing::AddressHasher;
2020-02-21 11:36:22 +01:00
use alloc::vec;
2020-02-17 15:28:10 +01:00
use alloc::vec::Vec;
use core::cmp::min;
use itertools::Itertools;
2020-02-17 15:28:10 +01:00
pub unsafe fn only_reload(p: *const u8) -> u64 {
2020-02-17 15:28:10 +01:00
let t = rdtsc_fence();
maccess(p);
2020-02-21 11:36:22 +01:00
rdtsc_fence() - t
2020-02-17 15:28:10 +01:00
}
pub unsafe fn flush_and_reload(p: *const u8) -> u64 {
2020-02-17 15:28:10 +01:00
flush(p);
let t = rdtsc_fence();
maccess(p);
2020-02-21 11:36:22 +01:00
rdtsc_fence() - t
2020-02-17 15:28:10 +01:00
}
2020-05-27 14:02:19 +02:00
pub unsafe fn only_flush(p: *const u8) -> u64 {
2020-02-18 08:45:15 +01:00
let t = rdtsc_fence();
flush(p);
2020-02-21 11:36:22 +01:00
rdtsc_fence() - t
2020-02-18 08:45:15 +01:00
}
2020-05-27 14:02:19 +02:00
pub unsafe fn load_and_flush(p: *const u8) -> u64 {
maccess(p);
2020-02-18 08:45:15 +01:00
let t = rdtsc_fence();
flush(p);
2020-02-21 11:36:22 +01:00
rdtsc_fence() - t
2020-02-18 08:45:15 +01:00
}
2020-05-27 14:02:19 +02:00
pub unsafe fn flush_and_flush(p: *const u8) -> u64 {
flush(p);
let t = rdtsc_fence();
flush(p);
rdtsc_fence() - t
}
2020-02-18 08:45:15 +01:00
pub unsafe fn l3_and_reload(p: *const u8) -> u64 {
flush(p);
arch_x86::_mm_mfence();
arch_x86::_mm_prefetch(p as *const i8, arch_x86::_MM_HINT_T2);
arch_x86::__cpuid_count(0, 0);
let t = rdtsc_fence();
maccess(p);
rdtsc_fence() - t
}
2020-02-18 08:45:15 +01:00
const BUCKET_SIZE: usize = 5;
const BUCKET_NUMBER: usize = 250;
// TODO same as below, also add the whole page calibration
2020-02-28 12:03:51 +01:00
pub fn calibrate_access(array: &[u8; 4096]) -> u64 {
println!("Calibrating...");
2020-02-17 15:28:10 +01:00
2020-02-18 08:45:15 +01:00
// Allocate a target array
// TBD why size, why the position in the array, why the type (usize)
2020-02-28 12:03:51 +01:00
// let mut array = Vec::<usize>::with_capacity(5 << 10);
// array.resize(5 << 10, 1);
2020-02-17 15:28:10 +01:00
2020-02-28 12:03:51 +01:00
// let array = array.into_boxed_slice();
2020-02-17 15:28:10 +01:00
2020-02-18 08:45:15 +01:00
// Histograms bucket of 5 and max at 400 cycles
// Magic numbers to be justified
// 80 is a size of screen
2020-02-21 11:36:22 +01:00
let mut hit_histogram = vec![0; BUCKET_NUMBER]; //Vec::<u32>::with_capacity(BUCKET_NUMBER);
//hit_histogram.resize(BUCKET_NUMBER, 0);
2020-02-17 15:28:10 +01:00
let mut miss_histogram = hit_histogram.clone();
2020-02-18 08:45:15 +01:00
// the address in memory we are going to target
2020-02-28 12:03:51 +01:00
let pointer = &array[0] as *const u8;
println!("buffer start {:p}", pointer);
2020-02-28 12:03:51 +01:00
if pointer as usize & 0x3f != 0 {
panic!("not aligned nicely");
}
2020-02-17 15:28:10 +01:00
2020-02-18 08:45:15 +01:00
// do a large sample of accesses to a cached line
2020-02-17 15:28:10 +01:00
unsafe { maccess(pointer) };
2020-02-28 12:03:51 +01:00
for i in 0..(4 << 10) {
for _ in 0..(1 << 10) {
let d = unsafe { only_reload(pointer.offset(i & (!0x3f))) } as usize;
hit_histogram[min(BUCKET_NUMBER - 1, d / BUCKET_SIZE) as usize] += 1;
}
2020-02-17 15:28:10 +01:00
}
2020-02-18 08:45:15 +01:00
// do a large numer of accesses to uncached line
2020-02-17 15:28:10 +01:00
unsafe { flush(pointer) };
2020-02-28 12:03:51 +01:00
for i in 0..(4 << 10) {
for _ in 0..(1 << 10) {
let d = unsafe { flush_and_reload(pointer.offset(i & (!0x3f))) } as usize;
miss_histogram[min(BUCKET_NUMBER - 1, d / BUCKET_SIZE) as usize] += 1;
}
2020-02-17 15:28:10 +01:00
}
2020-02-18 08:45:15 +01:00
let mut hit_max = 0;
let mut hit_max_i = 0;
let mut miss_min_i = 0;
for i in 0..hit_histogram.len() {
println!(
2020-02-18 08:45:15 +01:00
"{:3}: {:10} {:10}",
i * BUCKET_SIZE,
hit_histogram[i],
miss_histogram[i]
);
if hit_max < hit_histogram[i] {
hit_max = hit_histogram[i];
hit_max_i = i;
}
if miss_histogram[i] > 3 /* Magic */ && miss_min_i == 0 {
miss_min_i = i
}
}
println!("Miss min {}", miss_min_i * BUCKET_SIZE);
println!("Max hit {}", hit_max_i * BUCKET_SIZE);
2020-02-18 08:45:15 +01:00
let mut min = u32::max_value();
let mut min_i = 0;
for i in hit_max_i..miss_min_i {
if min > hit_histogram[i] + miss_histogram[i] {
min = hit_histogram[i] + miss_histogram[i];
min_i = i;
}
}
2020-02-17 15:28:10 +01:00
println!("Threshold {}", min_i * BUCKET_SIZE);
println!("Calibration done.");
2020-02-18 08:45:15 +01:00
(min_i * BUCKET_SIZE) as u64
2020-02-17 15:28:10 +01:00
}
2020-02-18 08:45:15 +01:00
const CFLUSH_BUCKET_SIZE: usize = 1;
2020-03-18 14:30:16 +01:00
const CFLUSH_BUCKET_NUMBER: usize = 500;
2020-02-18 08:45:15 +01:00
const CFLUSH_NUM_ITER: u32 = 1 << 11;
pub fn calibrate_flush(
array: &[u8],
cache_line_size: usize,
verbose_level: Verbosity,
) -> Vec<CalibrateResult> {
let pointer = (&array[0]) as *const u8;
if pointer as usize & (cache_line_size - 1) != 0 {
panic!("not aligned nicely");
}
2020-02-18 08:45:15 +01:00
calibrate_impl_fixed_freq(
pointer,
cache_line_size,
array.len() as isize,
&[
CalibrateOperation {
op: load_and_flush,
name: "clflush_hit",
display_name: "clflush hit",
},
CalibrateOperation {
op: flush_and_flush,
name: "clflush_miss",
display_name: "clflush miss",
},
],
HistParams {
bucket_number: CFLUSH_BUCKET_NUMBER,
bucket_size: CFLUSH_BUCKET_SIZE,
iterations: CFLUSH_NUM_ITER,
},
verbose_level,
None,
)
}
2020-02-18 08:45:15 +01:00
#[derive(Debug)]
pub struct CalibrateResult {
offset: isize,
histogram: Vec<Vec<u32>>,
median: Vec<u64>,
min: Vec<u64>,
max: Vec<u64>,
}
2020-02-18 08:45:15 +01:00
pub struct CalibrateOperation<'a> {
pub op: unsafe fn(*const u8) -> u64,
pub name: &'a str,
pub display_name: &'a str,
}
pub unsafe fn calibrate(
p: *const u8,
increment: usize,
len: isize,
operations: &[CalibrateOperation],
buckets_num: usize,
bucket_size: usize,
num_iterations: u32,
verbosity_level: Verbosity,
) -> Vec<CalibrateResult> {
calibrate_impl_fixed_freq(
p,
increment,
len,
operations,
HistParams {
bucket_number: buckets_num,
bucket_size,
iterations: num_iterations,
},
verbosity_level,
None,
)
}
2020-02-18 08:45:15 +01:00
const SPURIOUS_THRESHOLD: u32 = 1;
fn calibrate_impl_fixed_freq(
p: *const u8,
increment: usize,
len: isize,
operations: &[CalibrateOperation],
hist_params: HistParams,
verbosity_level: Verbosity,
hasher: Option<&AddressHasher>,
) -> Vec<CalibrateResult> {
// TODO : adapt this to detect CPU generation and grab the correct masks.
// These are the skylake masks.
/*let masks: [usize; 3] = [
0b1111_0011_0011_0011_0010_0100_1100_0100_000000,
0b1011_1010_1101_0111_1110_1010_1010_0010_000000,
0b0110_1101_0111_1101_0101_1101_0101_0001_000000,
];
let hasher = AddressHasher::new(&masks);
*/
if verbosity_level >= Thresholds {
println!(
"Calibrating {}...",
operations
.iter()
.map(|operation| { operation.display_name })
.format(", ")
);
2020-02-18 08:45:15 +01:00
}
2020-02-28 12:03:51 +01:00
let to_bucket = |time: u64| -> usize { time as usize / hist_params.bucket_size };
let from_bucket = |bucket: usize| -> u64 { (bucket * hist_params.bucket_size) as u64 };
let mut ret = Vec::new();
if verbosity_level >= Thresholds {
print!("CSV: address, ");
if hasher.is_some() {
print!("hash, ");
}
println!(
"{} min, {} median, {} max",
operations
.iter()
.map(|operation| operation.name)
.format(" min, "),
operations
.iter()
.map(|operation| operation.name)
.format(" median, "),
operations
.iter()
.map(|operation| operation.name)
.format(" max, ")
);
}
if verbosity_level >= RawResult {
print!("RESULT:address,");
if hasher.is_some() {
print!("hash,");
}
println!(
"time,{}",
operations
.iter()
.map(|operation| operation.name)
.format(",")
);
}
for i in (0..len).step_by(increment) {
let pointer = unsafe { p.offset(i) };
let hash = hasher.map(|h| h.hash(pointer as usize));
if verbosity_level >= Thresholds {
print!("Calibration for {:p}", pointer);
if let Some(h) = hash {
print!(" (hash: {:x})", h)
}
println!();
}
// TODO add some useful impl to CalibrateResults
let mut calibrate_result = CalibrateResult {
offset: i,
histogram: Vec::new(),
median: vec![0; operations.len()],
min: vec![0; operations.len()],
max: vec![0; operations.len()],
};
calibrate_result.histogram.reserve(operations.len());
for op in operations {
let mut hist = vec![0; hist_params.bucket_number];
for _ in 0..hist_params.iterations {
let time = unsafe { (op.op)(pointer) };
let bucket = min(hist_params.bucket_number - 1, to_bucket(time));
hist[bucket] += 1;
}
calibrate_result.histogram.push(hist);
2020-02-28 12:03:51 +01:00
}
2020-02-18 08:45:15 +01:00
let mut sums = vec![0; operations.len()];
let median_thresholds: Vec<u32> = calibrate_result
.histogram
.iter()
.map(|h| (hist_params.iterations - h[hist_params.bucket_number - 1]) / 2)
.collect();
2020-02-18 17:07:59 +01:00
for j in 0..hist_params.bucket_number - 1 {
if verbosity_level >= RawResult {
print!("RESULT:{:p},", pointer);
if let Some(h) = hash {
print!("{:x},", h);
}
print!("{}", from_bucket(j));
2020-02-28 12:03:51 +01:00
}
// ignore the last bucket : spurious context switches etc.
for op in 0..operations.len() {
let hist = &calibrate_result.histogram[op][j];
let min = &mut calibrate_result.min[op];
let max = &mut calibrate_result.max[op];
let med = &mut calibrate_result.median[op];
let sum = &mut sums[op];
if verbosity_level >= RawResult {
print!(",{}", hist);
}
if *min == 0 {
// looking for min
if *hist > SPURIOUS_THRESHOLD {
*min = from_bucket(j);
}
} else if *hist > SPURIOUS_THRESHOLD {
*max = from_bucket(j);
}
if *med == 0 {
*sum += *hist;
if *sum >= median_thresholds[op] {
*med = from_bucket(j);
}
}
2020-02-28 12:03:51 +01:00
}
if verbosity_level >= RawResult {
println!();
}
}
if verbosity_level >= Thresholds {
for (j, op) in operations.iter().enumerate() {
println!(
"{}: min {}, median {}, max {}",
op.display_name,
calibrate_result.min[j],
calibrate_result.median[j],
calibrate_result.max[j]
);
2020-02-28 12:03:51 +01:00
}
print!("CSV: {:p}, ", pointer);
if let Some(h) = hash {
print!("{:x}, ", h)
}
println!(
"{}, {}, {}",
calibrate_result.min.iter().format(", "),
calibrate_result.median.iter().format(", "),
calibrate_result.max.iter().format(", ")
);
}
ret.push(calibrate_result);
2020-02-28 12:03:51 +01:00
}
ret
2020-02-17 15:28:10 +01:00
}
#[allow(non_snake_case)]
pub fn calibrate_L3_miss_hit(
array: &[u8],
cache_line_size: usize,
verbose_level: Verbosity,
) -> CalibrateResult {
if verbose_level > NoOutput {
println!("Calibrating L3 access...");
}
let pointer = (&array[0]) as *const u8;
let r = calibrate_impl_fixed_freq(
pointer,
cache_line_size,
array.len() as isize,
&[CalibrateOperation {
op: l3_and_reload,
name: "l3_hit",
display_name: "L3 hit",
}],
HistParams {
bucket_number: 512,
bucket_size: 2,
iterations: 1 << 11,
},
verbose_level,
None,
);
r.into_iter().next().unwrap()
}