gaspersic/solana
1
0
Fork 0
Code Issues Pull Requests Projects Releases Wiki Activity

1. Persist to blockstore less frequently;

Browse Source 
2. reduce alpha for EMA to 1 percent to have roughly 200 data points for estimatio
This commit is contained in:
Tao Zhu
2022-02-04 18:57:02 -06:00
committed by Tao Zhu
parent 6587dbfa47
commit 7aa1fb4e24
4 changed files with 232 additions and 176 deletions
Download Patch File Download Diff File Expand all files Collapse all files

80
runtime/src/cost_model.rs
View File

@ -96,17 +96,11 @@ impl CostModel {
tx_cost
}
pub fn upsert_instruction_cost(
&mut self,
program_key: &Pubkey,
cost: u64,
) -> Result<u64, &'static str> {
// update-or-insert op is always successful. However the result of upsert, eg the aggregated
// value, requires additional calculation, which should only be envoked when needed.
pub fn upsert_instruction_cost(&mut self, program_key: &Pubkey, cost: u64) {
self.instruction_execution_cost_table
.upsert(program_key, cost);
match self.instruction_execution_cost_table.get_cost(program_key) {
Some(cost) => Ok(cost),
None => Err("failed to upsert to ExecuteCostTable"),
}
}
pub fn find_instruction_cost(&self, program_key: &Pubkey) -> u64 {
@ -115,7 +109,7 @@ impl CostModel {
None => {
let default_value = self.instruction_execution_cost_table.get_default();
debug!(
"Program key {:?} does not have assigned cost, using default value {}",
"instruction {:?} does not have aggregated cost, using default {}",
program_key, default_value
);
default_value
@ -123,6 +117,10 @@ impl CostModel {
}
}
pub fn get_program_keys(&self) -> Vec<&Pubkey> {
self.instruction_execution_cost_table.get_program_keys()
}
fn get_signature_cost(&self, transaction: &SanitizedTransaction) -> u64 {
transaction.signatures().len() as u64 * SIGNATURE_COST
}
@ -246,6 +244,7 @@ mod tests {
transaction::Transaction,
},
std::{
collections::HashMap,
str::FromStr,
sync::{Arc, RwLock},
thread::{self, JoinHandle},
@ -269,13 +268,11 @@ mod tests {
let mut testee = CostModel::default();
let known_key = Pubkey::from_str("known11111111111111111111111111111111111111").unwrap();
testee.upsert_instruction_cost(&known_key, 100).unwrap();
testee.upsert_instruction_cost(&known_key, 100);
// find cost for known programs
assert_eq!(100, testee.find_instruction_cost(&known_key));
testee
.upsert_instruction_cost(&bpf_loader::id(), 1999)
.unwrap();
testee.upsert_instruction_cost(&bpf_loader::id(), 1999);
assert_eq!(1999, testee.find_instruction_cost(&bpf_loader::id()));
// unknown program is assigned with default cost
@ -287,6 +284,35 @@ mod tests {
);
}
#[test]
fn test_iterating_instruction_cost_by_program_keys() {
solana_logger::setup();
let mut testee = CostModel::default();
let mut test_key_and_cost = HashMap::<Pubkey, u64>::new();
(0u64..10u64).for_each(|n| {
test_key_and_cost.insert(Pubkey::new_unique(), n);
});
test_key_and_cost.iter().for_each(|(key, cost)| {
testee.upsert_instruction_cost(key, *cost);
info!("key {:?} cost {}", key, cost);
});
let keys = testee.get_program_keys();
// verify each key has pre-set value
keys.iter().for_each(|key| {
let expected_cost = test_key_and_cost.get(key).unwrap();
info!(
"check key {:?} expect {} find {}",
key,
expected_cost,
testee.find_instruction_cost(key)
);
assert_eq!(*expected_cost, testee.find_instruction_cost(key));
});
}
#[test]
fn test_cost_model_data_len_cost() {
let lamports = 0;
@ -351,9 +377,7 @@ mod tests {
let expected_cost = 8;
let mut testee = CostModel::default();
testee
.upsert_instruction_cost(&system_program::id(), expected_cost)
.unwrap();
testee.upsert_instruction_cost(&system_program::id(), expected_cost);
assert_eq!(
expected_cost,
testee.get_transaction_cost(&simple_transaction)
@ -381,9 +405,7 @@ mod tests {
let expected_cost = program_cost * 2;
let mut testee = CostModel::default();
testee
.upsert_instruction_cost(&system_program::id(), program_cost)
.unwrap();
testee.upsert_instruction_cost(&system_program::id(), program_cost);
assert_eq!(expected_cost, testee.get_transaction_cost(&tx));
}
@ -464,7 +486,7 @@ mod tests {
);
// insert instruction cost to table
assert!(cost_model.upsert_instruction_cost(&key1, cost1).is_ok());
cost_model.upsert_instruction_cost(&key1, cost1);
// now it is known insturction with known cost
assert_eq!(cost1, cost_model.find_instruction_cost(&key1));
@ -484,9 +506,7 @@ mod tests {
let expected_execution_cost = 8;
let mut cost_model = CostModel::default();
cost_model
.upsert_instruction_cost(&system_program::id(), expected_execution_cost)
.unwrap();
cost_model.upsert_instruction_cost(&system_program::id(), expected_execution_cost);
let tx_cost = cost_model.calculate_cost(&tx);
assert_eq!(expected_account_cost, tx_cost.write_lock_cost);
assert_eq!(expected_execution_cost, tx_cost.execution_cost);
@ -498,17 +518,17 @@ mod tests {
let key1 = Pubkey::new_unique();
let cost1 = 100;
let cost2 = 200;
// updated_cost = (mean + 2*std)
let updated_cost = 238;
// updated_cost = (mean + 2*std) of [100, 200] => 120.899
let updated_cost = 121;
let mut cost_model = CostModel::default();
// insert instruction cost to table
assert!(cost_model.upsert_instruction_cost(&key1, cost1).is_ok());
cost_model.upsert_instruction_cost(&key1, cost1);
assert_eq!(cost1, cost_model.find_instruction_cost(&key1));
// update instruction cost
assert!(cost_model.upsert_instruction_cost(&key1, cost2).is_ok());
cost_model.upsert_instruction_cost(&key1, cost2);
assert_eq!(updated_cost, cost_model.find_instruction_cost(&key1));
}
@ -550,8 +570,8 @@ mod tests {
if i == 5 {
thread::spawn(move || {
let mut cost_model = cost_model.write().unwrap();
assert!(cost_model.upsert_instruction_cost(&prog1, cost1).is_ok());
assert!(cost_model.upsert_instruction_cost(&prog2, cost2).is_ok());
cost_model.upsert_instruction_cost(&prog1, cost1);
cost_model.upsert_instruction_cost(&prog2, cost2);
})
} else {
thread::spawn(move || {

85
runtime/src/execute_cost_table.rs
View File

@ -4,7 +4,10 @@
/// When its capacity limit is reached, it prunes old and less-used programs
/// to make room for new ones.
use log::*;
use {solana_sdk::pubkey::Pubkey, std::collections::HashMap};
use {
solana_sdk::pubkey::Pubkey,
std::collections::{hash_map::Entry, HashMap},
};
// prune is rather expensive op, free up bulk space in each operation
// would be more efficient. PRUNE_RATIO defines the after prune table
@ -18,7 +21,8 @@ const DEFAULT_CAPACITY: usize = 1024;
// The coefficient represents the degree of weighting decrease in EMA,
// a constant smoothing factor between 0 and 1. A higher alpha
// discounts older observations faster.
const COEFFICIENT: f64 = 0.4;
// Setting it using `2/(N+1)` where N is 200 samples
const COEFFICIENT: f64 = 0.01;
#[derive(Debug, Default)]
struct AggregatedVarianceStats {
@ -53,19 +57,27 @@ impl ExecuteCostTable {
self.table.len()
}
// default prorgam cost to max
// default program cost to max
pub fn get_default(&self) -> u64 {
// default max comoute units per program
// default max compute units per program
200_000u64
}
// returns None if program doesn't exist in table. In this case,
// it is advised to call `get_default()` for default program costdefault/
// it is advised to call `get_default()` for default program cost.
// Program cost is estimated as 2 standard deviations above mean, eg
// cost = (mean + 2 * std)
pub fn get_cost(&self, key: &Pubkey) -> Option<u64> {
let aggregated = self.table.get(key)?;
Some((aggregated.ema + 2.0 * aggregated.ema_var.sqrt()).ceil() as u64)
let cost_f64 = (aggregated.ema + 2.0 * aggregated.ema_var.sqrt()).ceil();
// check if cost:f64 can be losslessly convert to u64, otherwise return None
let cost_u64 = cost_f64 as u64;
if cost_f64 == cost_u64 as f64 {
Some(cost_u64)
} else {
None
}
}
pub fn upsert(&mut self, key: &Pubkey, value: u64) {
@ -77,21 +89,21 @@ impl ExecuteCostTable {
// exponential moving average algorithm
// https://en.wikipedia.org/wiki/Moving_average#Exponentially_weighted_moving_variance_and_standard_deviation
if self.table.contains_key(key) {
let aggregated = self.table.get_mut(key).unwrap();
let theta = value as f64 - aggregated.ema;
aggregated.ema += theta * COEFFICIENT;
aggregated.ema_var =
(1.0 - COEFFICIENT) * (aggregated.ema_var + COEFFICIENT * theta * theta)
} else {
// the starting values
self.table.insert(
*key,
AggregatedVarianceStats {
match self.table.entry(*key) {
Entry::Occupied(mut entry) => {
let aggregated = entry.get_mut();
let theta = value as f64 - aggregated.ema;
aggregated.ema += theta * COEFFICIENT;
aggregated.ema_var =
(1.0 - COEFFICIENT) * (aggregated.ema_var + COEFFICIENT * theta * theta);
}
Entry::Vacant(entry) => {
// the starting values
entry.insert(AggregatedVarianceStats {
ema: value as f64,
ema_var: 0.0,
},
);
});
}
}
let (count, timestamp) = self
@ -102,6 +114,10 @@ impl ExecuteCostTable {
*timestamp = Self::micros_since_epoch();
}
pub fn get_program_keys(&self) -> Vec<&Pubkey> {
self.table.keys().collect()
}
// prune the old programs so the table contains `new_size` of records,
// where `old` is defined as weighted age, which is negatively correlated
// with program's age and
@ -189,9 +205,9 @@ mod tests {
let key2 = Pubkey::new_unique();
let key3 = Pubkey::new_unique();
// simulate a lot of occurences to key1, so even there're longer than
// simulate a lot of occurrences to key1, so even there're longer than
// usual delay between upsert(key1..) and upsert(key2, ..), test
// would still satisfy as key1 has enough occurences to compensate
// would still satisfy as key1 has enough occurrences to compensate
// its age.
for i in 0..1000 {
testee.upsert(&key1, i);
@ -235,8 +251,8 @@ mod tests {
// update 1st record
testee.upsert(&key1, cost2);
assert_eq!(2, testee.get_count());
// expected key1 cost = (mean + 2*std) = (105 + 2*5) = 115
let expected_cost = 114;
// expected key1 cost is EMA of [100, 110] with alpha=0.01 => 103
let expected_cost = 103;
assert_eq!(expected_cost, testee.get_cost(&key1).unwrap());
assert_eq!(cost2, testee.get_cost(&key2).unwrap());
}
@ -280,10 +296,29 @@ mod tests {
testee.upsert(&key4, cost4);
assert_eq!(2, testee.get_count());
assert!(testee.get_cost(&key1).is_none());
// expected key2 cost = (mean + 2*std) = (105 + 2*5) = 115
let expected_cost_2 = 116;
// expected key2 cost = (mean + 2*std) of [110, 100] => 112
let expected_cost_2 = 112;
assert_eq!(expected_cost_2, testee.get_cost(&key2).unwrap());
assert!(testee.get_cost(&key3).is_none());
assert_eq!(cost4, testee.get_cost(&key4).unwrap());
}
#[test]
fn test_get_cost_overflow_u64() {
solana_logger::setup();
let mut testee = ExecuteCostTable::default();
let key1 = Pubkey::new_unique();
let cost1: u64 = f64::MAX as u64;
let cost2: u64 = u64::MAX / 2; // create large variance so the final result will overflow
// insert one record
testee.upsert(&key1, cost1);
assert_eq!(1, testee.get_count());
assert_eq!(cost1, testee.get_cost(&key1).unwrap());
// update cost
testee.upsert(&key1, cost2);
assert!(testee.get_cost(&key1).is_none());
}
}