Deuxfleurs/garage
41
56
Fork 73
Code Issues 102 Pull requests 14 Releases 30 Wiki Activity

Garage v0.9 #473

Merged
lx merged 175 commits from next into main 2023-10-10 13:28:29 +00:00
Conversation 0 Commits 175 Files changed 120 +795 -745
2 changed files with 795 additions and 745 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files
Showing only changes of commit 2aeaddd5e2 - Show all commits

336
src/rpc/layout.rs
View file

@ -1,12 +1,12 @@
use std::cmp::min;
use std::cmp::Ordering; use std::cmp::Ordering;
use std::cmp::{min}; use std::collections::HashMap;
use std::collections::{HashMap};
use serde::{Deserialize, Serialize}; use serde::{Deserialize, Serialize};
use garage_util::bipartite::*;
use garage_util::crdt::{AutoCrdt, Crdt, LwwMap}; use garage_util::crdt::{AutoCrdt, Crdt, LwwMap};
use garage_util::data::*; use garage_util::data::*;
use garage_util::bipartite::*;
use rand::prelude::SliceRandom; use rand::prelude::SliceRandom;
@ -168,7 +168,6 @@ impl ClusterLayout {
true true
} }
/// This function calculates a new partition-to-node assignation. /// This function calculates a new partition-to-node assignation.
/// The computed assignation maximizes the capacity of a /// The computed assignation maximizes the capacity of a
/// partition (assuming all partitions have the same size). /// partition (assuming all partitions have the same size).
@ -177,7 +176,6 @@ impl ClusterLayout {
/// data to be moved. A heuristic ensures node triplets /// data to be moved. A heuristic ensures node triplets
/// dispersion (in garage_util::bipartite::optimize_matching()). /// dispersion (in garage_util::bipartite::optimize_matching()).
pub fn calculate_partition_assignation(&mut self) -> bool { pub fn calculate_partition_assignation(&mut self) -> bool {
//The nodes might have been updated, some might have been deleted. //The nodes might have been updated, some might have been deleted.
//So we need to first update the list of nodes and retrieve the //So we need to first update the list of nodes and retrieve the
//assignation. //assignation.
@ -191,7 +189,8 @@ impl ClusterLayout {
//We collect part_per_zone in a vec to not rely on the //We collect part_per_zone in a vec to not rely on the
//arbitrary order in which elements are iterated in //arbitrary order in which elements are iterated in
//Hashmap::iter() //Hashmap::iter()
let part_per_zone_vec = part_per_zone.iter() let part_per_zone_vec = part_per_zone
.iter()
.map(|(x, y)| (x.clone(), *y)) .map(|(x, y)| (x.clone(), *y))
.collect::<Vec<(String, usize)>>(); .collect::<Vec<(String, usize)>>();
//We create an indexing of the zones //We create an indexing of the zones
@ -206,15 +205,11 @@ impl ClusterLayout {
let nb_nodes = part_per_nod.len(); let nb_nodes = part_per_nod.len();
let nb_partitions = 1 << PARTITION_BITS; let nb_partitions = 1 << PARTITION_BITS;
let left_cap_vec = vec![self.replication_factor as u32; nb_partitions]; let left_cap_vec = vec![self.replication_factor as u32; nb_partitions];
let right_cap_vec = part_per_zone_vec.iter().map(|(_,y)| *y as u32) let right_cap_vec = part_per_zone_vec.iter().map(|(_, y)| *y as u32).collect();
.collect(); let mut zone_assignation = dinic_compute_matching(left_cap_vec, right_cap_vec);
let mut zone_assignation =
dinic_compute_matching(left_cap_vec, right_cap_vec);
//We create the structure for the partition-to-node assignation. //We create the structure for the partition-to-node assignation.
let mut node_assignation = let mut node_assignation = vec![vec![None; self.replication_factor]; nb_partitions];
vec![vec![None; self.replication_factor ];nb_partitions];
//We will decrement part_per_nod to keep track of the number //We will decrement part_per_nod to keep track of the number
//of partitions that we still have to associate. //of partitions that we still have to associate.
let mut part_per_nod = part_per_nod.clone(); let mut part_per_nod = part_per_nod.clone();
@ -224,27 +219,33 @@ impl ClusterLayout {
//We get the id of the zones of the former assignation //We get the id of the zones of the former assignation
//(and the id no_zone if there is no node assignated) //(and the id no_zone if there is no node assignated)
let no_zone = part_per_zone_vec.len(); let no_zone = part_per_zone_vec.len();
let old_zone_assignation : Vec<Vec<usize>> = let old_zone_assignation: Vec<Vec<usize>> = old_node_assignation
old_node_assignation.iter().map(|x| x.iter().map( .iter()
|id| match *id { Some(i) => zone_id[&node_zone[i]] , .map(|x| {
None => no_zone } x.iter()
).collect()).collect(); .map(|id| match *id {
Some(i) => zone_id[&node_zone[i]],
None => no_zone,
})
.collect()
})
.collect();
//We minimize the distance to the former zone assignation //We minimize the distance to the former zone assignation
zone_assignation = optimize_matching( zone_assignation =
&old_zone_assignation, &zone_assignation, nb_zones+1); //+1 for no_zone optimize_matching(&old_zone_assignation, &zone_assignation, nb_zones + 1); //+1 for no_zone
//We need to assign partitions to nodes in their zone //We need to assign partitions to nodes in their zone
//We first put the nodes assignation that can stay the same //We first put the nodes assignation that can stay the same
for i in 0..nb_partitions { for i in 0..nb_partitions {
for j in 0..self.replication_factor { for j in 0..self.replication_factor {
if let Some(Some(former_node)) = old_node_assignation[i].iter().find( if let Some(Some(former_node)) = old_node_assignation[i].iter().find(|x| {
|x| if let Some(id) = x { if let Some(id) = x {
zone_id[&node_zone[*id]] == zone_assignation[i][j] zone_id[&node_zone[*id]] == zone_assignation[i][j]
} else {
false
} }
else {false} }) {
)
{
if part_per_nod[*former_node] > 0 { if part_per_nod[*former_node] > 0 {
node_assignation[i][j] = Some(*former_node); node_assignation[i][j] = Some(*former_node);
part_per_nod[*former_node] -= 1; part_per_nod[*former_node] -= 1;
@ -253,16 +254,17 @@ impl ClusterLayout {
} }
} }
//We complete the assignation of partitions to nodes //We complete the assignation of partitions to nodes
let mut rng = rand::thread_rng(); let mut rng = rand::thread_rng();
for i in 0..nb_partitions { for i in 0..nb_partitions {
for j in 0..self.replication_factor { for j in 0..self.replication_factor {
if node_assignation[i][j] == None { if node_assignation[i][j] == None {
let possible_nodes: Vec<usize> = (0..nb_nodes) let possible_nodes: Vec<usize> = (0..nb_nodes)
.filter( .filter(|id| {
|id| zone_id[&node_zone[*id]] == zone_assignation[i][j] zone_id[&node_zone[*id]] == zone_assignation[i][j]
&& part_per_nod[*id] > 0).collect(); && part_per_nod[*id] > 0
})
.collect();
assert!(possible_nodes.len() > 0); assert!(possible_nodes.len() > 0);
//We randomly pick a node //We randomly pick a node
if let Some(nod) = possible_nodes.choose(&mut rng) { if let Some(nod) = possible_nodes.choose(&mut rng) {
@ -279,14 +281,16 @@ impl ClusterLayout {
for j in 0..self.replication_factor { for j in 0..self.replication_factor {
if let Some(id) = node_assignation[i][j] { if let Some(id) = node_assignation[i][j] {
self.ring_assignation_data.push(id as CompactNodeType); self.ring_assignation_data.push(id as CompactNodeType);
} else {
assert!(false)
} }
else {assert!(false)}
} }
} }
true true
} else {
false
} }
else { false }
} }
/// The LwwMap of node roles might have changed. This function updates the node_id_vec /// The LwwMap of node roles might have changed. This function updates the node_id_vec
@ -296,19 +300,17 @@ impl ClusterLayout {
/// do modify assignation_ring and node_id_vec. /// do modify assignation_ring and node_id_vec.
fn update_nodes_and_ring(&mut self) -> Vec<Vec<Option<usize>>> { fn update_nodes_and_ring(&mut self) -> Vec<Vec<Option<usize>>> {
let nb_partitions = 1usize << PARTITION_BITS; let nb_partitions = 1usize << PARTITION_BITS;
let mut node_assignation = let mut node_assignation = vec![vec![None; self.replication_factor]; nb_partitions];
vec![vec![None; self.replication_factor ];nb_partitions];
let rf = self.replication_factor; let rf = self.replication_factor;
let ring = &self.ring_assignation_data; let ring = &self.ring_assignation_data;
let new_node_id_vec : Vec::<Uuid> = self.roles.items().iter() let new_node_id_vec: Vec<Uuid> = self.roles.items().iter().map(|(k, _, _)| *k).collect();
.map(|(k, _, _)| *k)
.collect();
if ring.len() == rf * nb_partitions { if ring.len() == rf * nb_partitions {
for i in 0..nb_partitions { for i in 0..nb_partitions {
for j in 0..self.replication_factor { for j in 0..self.replication_factor {
node_assignation[i][j] = new_node_id_vec.iter() node_assignation[i][j] = new_node_id_vec
.iter()
.position(|id| *id == self.node_id_vec[ring[i * rf + j] as usize]); .position(|id| *id == self.node_id_vec[ring[i * rf + j] as usize]);
} }
} }
@ -324,18 +326,18 @@ impl ClusterLayout {
///self.replication_factor times and the capacity of a partition ///self.replication_factor times and the capacity of a partition
///is maximized. ///is maximized.
fn optimal_proportions(&mut self) -> Option<(Vec<usize>, HashMap<String, usize>)> { fn optimal_proportions(&mut self) -> Option<(Vec<usize>, HashMap<String, usize>)> {
let mut zone_capacity: HashMap<String, u32> = HashMap::new(); let mut zone_capacity: HashMap<String, u32> = HashMap::new();
let (node_zone, node_capacity) = self.get_node_zone_capacity(); let (node_zone, node_capacity) = self.get_node_zone_capacity();
let nb_nodes = self.node_id_vec.len(); let nb_nodes = self.node_id_vec.len();
for i in 0..nb_nodes for i in 0..nb_nodes {
{
if zone_capacity.contains_key(&node_zone[i]) { if zone_capacity.contains_key(&node_zone[i]) {
zone_capacity.insert(node_zone[i].clone(), zone_capacity[&node_zone[i]] + node_capacity[i]); zone_capacity.insert(
} node_zone[i].clone(),
else{ zone_capacity[&node_zone[i]] + node_capacity[i],
);
} else {
zone_capacity.insert(node_zone[i].clone(), node_capacity[i]); zone_capacity.insert(node_zone[i].clone(), node_capacity[i]);
} }
} }
@ -355,37 +357,52 @@ impl ClusterLayout {
//However, a large zone can be associated to at most //However, a large zone can be associated to at most
//nb_partitions to ensure replication of the date. //nb_partitions to ensure replication of the date.
//So we take the min with nb_partitions: //So we take the min with nb_partitions:
let mut part_per_zone : HashMap<String, usize> = let mut part_per_zone: HashMap<String, usize> = zone_capacity
zone_capacity.iter() .iter()
.map(|(k, v)| (k.clone(), min(nb_partitions, .map(|(k, v)| {
(self.replication_factor*nb_partitions (
**v as usize)/sum_capacities as usize) ) ).collect(); k.clone(),
min(
nb_partitions,
(self.replication_factor * nb_partitions * *v as usize)
/ sum_capacities as usize,
),
)
})
.collect();
//The replication_factor-1 upper bounds the number of //The replication_factor-1 upper bounds the number of
//part_per_zones that are greater than nb_partitions //part_per_zones that are greater than nb_partitions
for _ in 1..self.replication_factor { for _ in 1..self.replication_factor {
//The number of partitions that are not assignated to //The number of partitions that are not assignated to
//a zone that takes nb_partitions. //a zone that takes nb_partitions.
let sum_capleft : u32 = zone_capacity.keys() let sum_capleft: u32 = zone_capacity
.filter(| k | {part_per_zone[*k] < nb_partitions} ) .keys()
.map(|k| zone_capacity[k]).sum(); .filter(|k| part_per_zone[*k] < nb_partitions)
.map(|k| zone_capacity[k])
.sum();
//The number of replication of the data that we need //The number of replication of the data that we need
//to ensure. //to ensure.
let repl_left = self.replication_factor let repl_left = self.replication_factor
- part_per_zone.values() - part_per_zone
.filter(|x| {**x == nb_partitions}) .values()
.filter(|x| **x == nb_partitions)
.count(); .count();
if repl_left == 0 { if repl_left == 0 {
break; break;
} }
for k in zone_capacity.keys() { for k in zone_capacity.keys() {
if part_per_zone[k] != nb_partitions if part_per_zone[k] != nb_partitions {
{ part_per_zone.insert(
part_per_zone.insert(k.to_string() , min(nb_partitions, k.to_string(),
(nb_partitions*zone_capacity[k] as usize min(
*repl_left)/sum_capleft as usize)); nb_partitions,
(nb_partitions * zone_capacity[k] as usize * repl_left)
/ sum_capleft as usize,
),
);
} }
} }
} }
@ -393,20 +410,21 @@ impl ClusterLayout {
//Now we divide the zone's partition share proportionally //Now we divide the zone's partition share proportionally
//between their nodes. //between their nodes.
let mut part_per_nod : Vec<usize> = (0..nb_nodes).map( let mut part_per_nod: Vec<usize> = (0..nb_nodes)
|i| (part_per_zone[&node_zone[i]]*node_capacity[i] as usize)/zone_capacity[&node_zone[i]] as usize .map(|i| {
) (part_per_zone[&node_zone[i]] * node_capacity[i] as usize)
/ zone_capacity[&node_zone[i]] as usize
})
.collect(); .collect();
//We must update the part_per_zone to make it correspond to //We must update the part_per_zone to make it correspond to
//part_per_nod (because of integer rounding) //part_per_nod (because of integer rounding)
part_per_zone = part_per_zone.iter().map(|(k,_)| part_per_zone = part_per_zone.iter().map(|(k, _)| (k.clone(), 0)).collect();
(k.clone(), 0))
.collect();
for i in 0..nb_nodes { for i in 0..nb_nodes {
part_per_zone.insert( part_per_zone.insert(
node_zone[i].clone(), node_zone[i].clone(),
part_per_zone[&node_zone[i]] + part_per_nod[i]); part_per_zone[&node_zone[i]] + part_per_nod[i],
);
} }
//Because of integer rounding, the total sum of part_per_nod //Because of integer rounding, the total sum of part_per_nod
@ -419,8 +437,7 @@ impl ClusterLayout {
// part_per_zone capped // part_per_zone capped
let discrepancy: usize = let discrepancy: usize =
nb_partitions*self.replication_factor nb_partitions * self.replication_factor - part_per_nod.iter().sum::<usize>();
- part_per_nod.iter().sum::<usize>();
//We use a stupid O(N^2) algorithm. If the number of nodes //We use a stupid O(N^2) algorithm. If the number of nodes
//is actually expected to be high, one should optimize this. //is actually expected to be high, one should optimize this.
@ -428,64 +445,77 @@ impl ClusterLayout {
for _ in 0..discrepancy { for _ in 0..discrepancy {
if let Some(idmax) = (0..nb_nodes) if let Some(idmax) = (0..nb_nodes)
.filter(|i| part_per_zone[&node_zone[*i]] < nb_partitions) .filter(|i| part_per_zone[&node_zone[*i]] < nb_partitions)
.max_by( |i,j| .max_by(|i, j| {
(node_capacity[*i] * (part_per_nod[*j] + 1) as u32) (node_capacity[*i] * (part_per_nod[*j] + 1) as u32)
.cmp(&(node_capacity[*j] * (part_per_nod[*i] + 1) as u32)) .cmp(&(node_capacity[*j] * (part_per_nod[*i] + 1) as u32))
) }) {
{
part_per_nod[idmax] += 1; part_per_nod[idmax] += 1;
part_per_zone.insert(node_zone[idmax].clone(),part_per_zone[&node_zone[idmax]]+1); part_per_zone.insert(
node_zone[idmax].clone(),
part_per_zone[&node_zone[idmax]] + 1,
);
} }
} }
//We check the algorithm consistency //We check the algorithm consistency
let discrepancy: usize = let discrepancy: usize =
nb_partitions*self.replication_factor nb_partitions * self.replication_factor - part_per_nod.iter().sum::<usize>();
- part_per_nod.iter().sum::<usize>();
assert!(discrepancy == 0); assert!(discrepancy == 0);
assert!(if let Some(v) = part_per_zone.values().max() assert!(if let Some(v) = part_per_zone.values().max() {
{*v <= nb_partitions} else {false} ); *v <= nb_partitions
} else {
false
});
Some((part_per_nod, part_per_zone)) Some((part_per_nod, part_per_zone))
} }
//Returns vectors of zone and capacity; indexed by the same (temporary) //Returns vectors of zone and capacity; indexed by the same (temporary)
//indices as node_id_vec. //indices as node_id_vec.
fn get_node_zone_capacity(&self) -> (Vec<String>, Vec<u32>) { fn get_node_zone_capacity(&self) -> (Vec<String>, Vec<u32>) {
let node_zone = self
.node_id_vec
.iter()
.map(|id_nod| match self.node_role(id_nod) {
Some(NodeRole {
zone,
capacity: _,
tags: _,
}) => zone.clone(),
_ => "".to_string(),
})
.collect();
let node_zone = self.node_id_vec.iter().map( let node_capacity = self
|id_nod| match self.node_role(id_nod) { .node_id_vec
Some(NodeRole{zone,capacity:_,tags:_}) => zone.clone() , .iter()
_ => "".to_string() .map(|id_nod| match self.node_role(id_nod) {
Some(NodeRole {
zone: _,
capacity,
tags: _,
}) => {
if let Some(c) = capacity {
*c
} else {
0
} }
).collect();
let node_capacity = self.node_id_vec.iter().map(
|id_nod| match self.node_role(id_nod) {
Some(NodeRole{zone:_,capacity,tags:_}) =>
if let Some(c)=capacity
{*c}
else {0},
_ => 0
} }
).collect(); _ => 0,
})
.collect();
(node_zone, node_capacity) (node_zone, node_capacity)
} }
} }
#[cfg(test)] #[cfg(test)]
mod tests { mod tests {
use super::*; use super::*;
use itertools::Itertools; use itertools::Itertools;
fn check_assignation(cl: &ClusterLayout) { fn check_assignation(cl: &ClusterLayout) {
//Check that input data has the right format //Check that input data has the right format
let nb_partitions = 1usize << PARTITION_BITS; let nb_partitions = 1usize << PARTITION_BITS;
assert!([1, 2, 3].contains(&cl.replication_factor)); assert!([1, 2, 3].contains(&cl.replication_factor));
@ -493,93 +523,117 @@ mod tests {
let (node_zone, node_capacity) = cl.get_node_zone_capacity(); let (node_zone, node_capacity) = cl.get_node_zone_capacity();
//Check that is is a correct assignation with zone redundancy //Check that is is a correct assignation with zone redundancy
let rf = cl.replication_factor; let rf = cl.replication_factor;
for i in 0..nb_partitions { for i in 0..nb_partitions {
assert!( rf == assert!(
cl.ring_assignation_data[rf*i..rf*(i+1)].iter() rf == cl.ring_assignation_data[rf * i..rf * (i + 1)]
.iter()
.map(|nod| node_zone[*nod as usize].clone()) .map(|nod| node_zone[*nod as usize].clone())
.unique() .unique()
.count() ); .count()
);
} }
let nb_nodes = cl.node_id_vec.len(); let nb_nodes = cl.node_id_vec.len();
//Check optimality //Check optimality
let node_nb_part =(0..nb_nodes).map(|i| cl.ring_assignation_data let node_nb_part = (0..nb_nodes)
.map(|i| {
cl.ring_assignation_data
.iter() .iter()
.filter(|x| **x == i as u8) .filter(|x| **x == i as u8)
.count()) .count()
.collect::<Vec::<_>>(); })
.collect::<Vec<_>>();
let zone_vec = node_zone.iter().unique().collect::<Vec::<_>>(); let zone_vec = node_zone.iter().unique().collect::<Vec<_>>();
let zone_nb_part = zone_vec.iter().map( |z| cl.ring_assignation_data.iter() let zone_nb_part = zone_vec
.iter()
.map(|z| {
cl.ring_assignation_data
.iter()
.filter(|x| node_zone[**x as usize] == **z) .filter(|x| node_zone[**x as usize] == **z)
.count() .count()
).collect::<Vec::<_>>(); })
.collect::<Vec<_>>();
//Check optimality of the zone assignation : would it be better for the //Check optimality of the zone assignation : would it be better for the
//node_capacity/node_partitions ratio to change the assignation of a partition //node_capacity/node_partitions ratio to change the assignation of a partition
if let Some(idmin) = (0..nb_nodes).min_by( if let Some(idmin) = (0..nb_nodes).min_by(|i, j| {
|i,j| (node_capacity[*i]*node_nb_part[*j] as u32) (node_capacity[*i] * node_nb_part[*j] as u32)
.cmp(&(node_capacity[*j] * node_nb_part[*i] as u32)) .cmp(&(node_capacity[*j] * node_nb_part[*i] as u32))
){ }) {
if let Some(idnew) = (0..nb_nodes) if let Some(idnew) = (0..nb_nodes)
.filter( |i| if let Some(p) = zone_vec.iter().position(|z| **z==node_zone[*i]) .filter(|i| {
{zone_nb_part[p] < nb_partitions } if let Some(p) = zone_vec.iter().position(|z| **z == node_zone[*i]) {
else { false }) zone_nb_part[p] < nb_partitions
.max_by( } else {
|i,j| (node_capacity[*i]*(node_nb_part[*j]as u32+1)) false
.cmp(&(node_capacity[*j]*(node_nb_part[*i] as u32+1))) }
){ })
assert!(node_capacity[idmin]*(node_nb_part[idnew] as u32+1) >= .max_by(|i, j| {
node_capacity[idnew]*node_nb_part[idmin] as u32); (node_capacity[*i] * (node_nb_part[*j] as u32 + 1))
.cmp(&(node_capacity[*j] * (node_nb_part[*i] as u32 + 1)))
}) {
assert!(
node_capacity[idmin] * (node_nb_part[idnew] as u32 + 1)
>= node_capacity[idnew] * node_nb_part[idmin] as u32
);
} }
} }
//In every zone, check optimality of the nod assignation //In every zone, check optimality of the nod assignation
for z in zone_vec { for z in zone_vec {
let node_of_z_iter = (0..nb_nodes).filter(|id| node_zone[*id] == *z); let node_of_z_iter = (0..nb_nodes).filter(|id| node_zone[*id] == *z);
if let Some(idmin) = node_of_z_iter.clone().min_by( if let Some(idmin) = node_of_z_iter.clone().min_by(|i, j| {
|i,j| (node_capacity[*i]*node_nb_part[*j] as u32) (node_capacity[*i] * node_nb_part[*j] as u32)
.cmp(&(node_capacity[*j] * node_nb_part[*i] as u32)) .cmp(&(node_capacity[*j] * node_nb_part[*i] as u32))
){ }) {
if let Some(idnew) = node_of_z_iter.min_by( if let Some(idnew) = node_of_z_iter.min_by(|i, j| {
|i,j| (node_capacity[*i]*(node_nb_part[*j] as u32+1)) (node_capacity[*i] * (node_nb_part[*j] as u32 + 1))
.cmp(&(node_capacity[*j] * (node_nb_part[*i] as u32 + 1))) .cmp(&(node_capacity[*j] * (node_nb_part[*i] as u32 + 1)))
}) {
assert!(
node_capacity[idmin] * (node_nb_part[idnew] as u32 + 1)
>= node_capacity[idnew] * node_nb_part[idmin] as u32
);
}
}
}
}
fn update_layout(
cl: &mut ClusterLayout,
node_id_vec: &Vec<u8>,
node_capacity_vec: &Vec<u32>,
node_zone_vec: &Vec<String>,
) { ) {
assert!(node_capacity[idmin]*(node_nb_part[idnew] as u32+1) >=
node_capacity[idnew]*node_nb_part[idmin] as u32);
}
}
}
}
fn update_layout(cl : &mut ClusterLayout, node_id_vec : &Vec<u8>,
node_capacity_vec : &Vec<u32> , node_zone_vec : &Vec<String>) {
for i in 0..node_id_vec.len() { for i in 0..node_id_vec.len() {
if let Some(x) = FixedBytes32::try_from(&[i as u8; 32]) { if let Some(x) = FixedBytes32::try_from(&[i as u8; 32]) {
cl.node_id_vec.push(x); cl.node_id_vec.push(x);
} }
let update = cl.roles.update_mutator(cl.node_id_vec[i] , let update = cl.roles.update_mutator(
cl.node_id_vec[i],
NodeRoleV(Some(NodeRole { NodeRoleV(Some(NodeRole {
zone: (node_zone_vec[i].to_string()), zone: (node_zone_vec[i].to_string()),
capacity: (Some(node_capacity_vec[i])), capacity: (Some(node_capacity_vec[i])),
tags : (vec![])}))); tags: (vec![]),
})),
);
cl.roles.merge(&update); cl.roles.merge(&update);
} }
} }
#[test] #[test]
fn test_assignation() { fn test_assignation() {
let mut node_id_vec = vec![1, 2, 3]; let mut node_id_vec = vec![1, 2, 3];
let mut node_capacity_vec = vec![4000, 1000, 2000]; let mut node_capacity_vec = vec![4000, 1000, 2000];
let mut node_zone_vec= vec!["A", "B", "C"].into_iter().map(|x| x.to_string()).collect(); let mut node_zone_vec = vec!["A", "B", "C"]
.into_iter()
.map(|x| x.to_string())
.collect();
let mut cl = ClusterLayout { let mut cl = ClusterLayout {
node_id_vec: vec![], node_id_vec: vec![],
@ -598,7 +652,10 @@ mod tests {
node_id_vec = vec![1, 2, 3, 4, 5, 6, 7, 8, 9]; node_id_vec = vec![1, 2, 3, 4, 5, 6, 7, 8, 9];
node_capacity_vec = vec![4000, 1000, 1000, 3000, 1000, 1000, 2000, 10000, 2000]; node_capacity_vec = vec![4000, 1000, 1000, 3000, 1000, 1000, 2000, 10000, 2000];
node_zone_vec= vec!["A", "B", "C", "C", "C", "B", "G", "H", "I"].into_iter().map(|x| x.to_string()).collect(); node_zone_vec = vec!["A", "B", "C", "C", "C", "B", "G", "H", "I"]
.into_iter()
.map(|x| x.to_string())
.collect();
update_layout(&mut cl, &node_id_vec, &node_capacity_vec, &node_zone_vec); update_layout(&mut cl, &node_id_vec, &node_capacity_vec, &node_zone_vec);
cl.calculate_partition_assignation(); cl.calculate_partition_assignation();
check_assignation(&cl); check_assignation(&cl);
@ -608,14 +665,9 @@ mod tests {
cl.calculate_partition_assignation(); cl.calculate_partition_assignation();
check_assignation(&cl); check_assignation(&cl);
node_capacity_vec = vec![4000, 4000, 2000, 7000, 1000, 9000, 2000, 10, 2000]; node_capacity_vec = vec![4000, 4000, 2000, 7000, 1000, 9000, 2000, 10, 2000];
update_layout(&mut cl, &node_id_vec, &node_capacity_vec, &node_zone_vec); update_layout(&mut cl, &node_id_vec, &node_capacity_vec, &node_zone_vec);
cl.calculate_partition_assignation(); cl.calculate_partition_assignation();
check_assignation(&cl); check_assignation(&cl);
} }
} }

90
src/util/bipartite.rs
View file

@ -4,9 +4,9 @@
* assignation. * assignation.
* */ * */
use std::cmp::{min,max};
use std::collections::VecDeque;
use rand::prelude::SliceRandom; use rand::prelude::SliceRandom;
use std::cmp::{max, min};
use std::collections::VecDeque;
//Graph data structure for the flow algorithm. //Graph data structure for the flow algorithm.
#[derive(Clone, Copy, Debug)] #[derive(Clone, Copy, Debug)]
@ -25,16 +25,16 @@ struct WeightedEdge{
v: usize, v: usize,
} }
/* This function takes two matchings (old_match and new_match) in a /* This function takes two matchings (old_match and new_match) in a
* complete bipartite graph. It returns a matching that has the * complete bipartite graph. It returns a matching that has the
* same degree as new_match at every vertex, and that is as close * same degree as new_match at every vertex, and that is as close
* as possible to old_match. * as possible to old_match.
* */ * */
pub fn optimize_matching( old_match : &Vec<Vec<usize>> , pub fn optimize_matching(
old_match: &Vec<Vec<usize>>,
new_match: &Vec<Vec<usize>>, new_match: &Vec<Vec<usize>>,
nb_right : usize ) nb_right: usize,
-> Vec<Vec<usize>> { ) -> Vec<Vec<usize>> {
let nb_left = old_match.len(); let nb_left = old_match.len();
let ed = WeightedEdge { w: -1, u: 0, v: 0 }; let ed = WeightedEdge { w: -1, u: 0, v: 0 };
let mut edge_vec = vec![ed; nb_left * nb_right]; let mut edge_vec = vec![ed; nb_left * nb_right];
@ -55,8 +55,7 @@ pub fn optimize_matching( old_match : &Vec<Vec<usize>> ,
} }
//We add the new matchings //We add the new matchings
if new_match[edge_vec[i].u].contains(&(edge_vec[i].v - nb_left)) { if new_match[edge_vec[i].u].contains(&(edge_vec[i].v - nb_left)) {
(edge_vec[i].u,edge_vec[i].v) = (edge_vec[i].u, edge_vec[i].v) = (edge_vec[i].v, edge_vec[i].u);
(edge_vec[i].v,edge_vec[i].u);
edge_vec[i].w *= -1; edge_vec[i].w *= -1;
} }
} }
@ -77,12 +76,10 @@ pub fn optimize_matching( old_match : &Vec<Vec<usize>> ,
if let Some(cycle) = positive_cycle(&edge_vec, nb_left, nb_right) { if let Some(cycle) = positive_cycle(&edge_vec, nb_left, nb_right) {
for i in cycle { for i in cycle {
//We flip the edges of the cycle. //We flip the edges of the cycle.
(edge_vec[i].u,edge_vec[i].v) = (edge_vec[i].u, edge_vec[i].v) = (edge_vec[i].v, edge_vec[i].u);
(edge_vec[i].v,edge_vec[i].u);
edge_vec[i].w *= -1; edge_vec[i].w *= -1;
} }
} } else {
else {
//If there is no cycle, we return the optimal matching. //If there is no cycle, we return the optimal matching.
break; break;
} }
@ -99,8 +96,11 @@ pub fn optimize_matching( old_match : &Vec<Vec<usize>> ,
} }
//This function finds a positive cycle in a bipartite wieghted graph. //This function finds a positive cycle in a bipartite wieghted graph.
fn positive_cycle( edge_vec : &Vec<WeightedEdge>, nb_left : usize, fn positive_cycle(
nb_right : usize) -> Option<Vec<usize>> { edge_vec: &Vec<WeightedEdge>,
nb_left: usize,
nb_right: usize,
) -> Option<Vec<usize>> {
let nb_side_min = min(nb_left, nb_right); let nb_side_min = min(nb_left, nb_right);
let nb_vertices = nb_left + nb_right; let nb_vertices = nb_left + nb_right;
let weight_lowerbound = -((nb_left + nb_right) as i32) - 1; let weight_lowerbound = -((nb_left + nb_right) as i32) - 1;
@ -164,22 +164,23 @@ fn positive_cycle( edge_vec : &Vec<WeightedEdge>, nb_left : usize,
None None
} }
// This function takes two arrays of capacity and computes the // This function takes two arrays of capacity and computes the
// maximal matching in the complete bipartite graph such that the // maximal matching in the complete bipartite graph such that the
// left vertex i is matched to left_cap_vec[i] right vertices, and // left vertex i is matched to left_cap_vec[i] right vertices, and
// the right vertex j is matched to right_cap_vec[j] left vertices. // the right vertex j is matched to right_cap_vec[j] left vertices.
// To do so, we use Dinic's maximum flow algorithm. // To do so, we use Dinic's maximum flow algorithm.
pub fn dinic_compute_matching( left_cap_vec : Vec<u32>, pub fn dinic_compute_matching(left_cap_vec: Vec<u32>, right_cap_vec: Vec<u32>) -> Vec<Vec<usize>> {
right_cap_vec : Vec<u32>) -> Vec< Vec<usize> > let mut graph = Vec::<Vec<EdgeFlow>>::new();
{ let ed = EdgeFlow {
let mut graph = Vec::<Vec::<EdgeFlow> >::new(); c: 0,
let ed = EdgeFlow{c:0,flow:0,v:0, rev:0}; flow: 0,
v: 0,
rev: 0,
};
// 0 will be the source // 0 will be the source
graph.push(vec![ed; left_cap_vec.len()]); graph.push(vec![ed; left_cap_vec.len()]);
for i in 0..left_cap_vec.len() for i in 0..left_cap_vec.len() {
{
graph[0][i].c = left_cap_vec[i] as i32; graph[0][i].c = left_cap_vec[i] as i32;
graph[0][i].v = i + 2; graph[0][i].v = i + 2;
graph[0][i].rev = 0; graph[0][i].rev = 0;
@ -187,8 +188,7 @@ pub fn dinic_compute_matching( left_cap_vec : Vec<u32>,
//1 will be the sink //1 will be the sink
graph.push(vec![ed; right_cap_vec.len()]); graph.push(vec![ed; right_cap_vec.len()]);
for i in 0..right_cap_vec.len() for i in 0..right_cap_vec.len() {
{
graph[1][i].c = right_cap_vec[i] as i32; graph[1][i].c = right_cap_vec[i] as i32;
graph[1][i].v = i + 2 + left_cap_vec.len(); graph[1][i].v = i + 2 + left_cap_vec.len();
graph[1][i].rev = 0; graph[1][i].rev = 0;
@ -270,16 +270,14 @@ pub fn dinic_compute_matching( left_cap_vec : Vec<u32>,
let flow_upper_bound; let flow_upper_bound;
if let Some(x) = left_cap_vec.iter().max() { if let Some(x) = left_cap_vec.iter().max() {
flow_upper_bound = *x as i32; flow_upper_bound = *x as i32;
} } else {
else {
flow_upper_bound = 0; flow_upper_bound = 0;
assert!(false); assert!(false);
} }
lifo.push_back((0, flow_upper_bound)); lifo.push_back((0, flow_upper_bound));
loop loop {
{
if let Some((id_tmp, f_tmp)) = lifo.back() { if let Some((id_tmp, f_tmp)) = lifo.back() {
let id = *id_tmp; let id = *id_tmp;
let f = *f_tmp; let f = *f_tmp;
@ -310,29 +308,30 @@ pub fn dinic_compute_matching( left_cap_vec : Vec<u32>,
continue; continue;
} }
//else we can try to send flow from id to its nbd //else we can try to send flow from id to its nbd
let new_flow = min(f,graph[id][nbd].c let new_flow = min(f, graph[id][nbd].c - graph[id][nbd].flow);
- graph[id][nbd].flow); if level[graph[id][nbd].v] <= level[id] || new_flow == 0 {
if level[graph[id][nbd].v] <= level[id] ||
new_flow == 0 {
//We cannot send flow to nbd. //We cannot send flow to nbd.
next_nbd[id] += 1; next_nbd[id] += 1;
continue; continue;
} }
//otherwise, we send flow to nbd. //otherwise, we send flow to nbd.
lifo.push_back((graph[id][nbd].v, new_flow)); lifo.push_back((graph[id][nbd].v, new_flow));
} } else {
else {
break; break;
} }
} }
} }
//We return the association //We return the association
let assoc_table = (0..left_cap_vec.len()).map( let assoc_table = (0..left_cap_vec.len())
|id| graph[id+2].iter() .map(|id| {
graph[id + 2]
.iter()
.filter(|e| e.flow > 0) .filter(|e| e.flow > 0)
.map(|e| e.v - 2 - left_cap_vec.len()) .map(|e| e.v - 2 - left_cap_vec.len())
.collect()).collect(); .collect()
})
.collect();
//consistency check //consistency check
@ -346,20 +345,21 @@ pub fn dinic_compute_matching( left_cap_vec : Vec<u32>,
//it solves the matching problem //it solves the matching problem
for i in 0..left_cap_vec.len() { for i in 0..left_cap_vec.len() {
assert!(left_cap_vec[i] as i32 == assert!(left_cap_vec[i] as i32 == graph[i + 2].iter().map(|e| max(0, e.flow)).sum::<i32>());
graph[i+2].iter().map(|e| max(0,e.flow)).sum::<i32>());
} }
for i in 0..right_cap_vec.len() { for i in 0..right_cap_vec.len() {
assert!(right_cap_vec[i] as i32 == assert!(
graph[i+2+left_cap_vec.len()].iter() right_cap_vec[i] as i32
.map(|e| max(0,e.flow)).sum::<i32>()); == graph[i + 2 + left_cap_vec.len()]
.iter()
.map(|e| max(0, e.flow))
.sum::<i32>()
);
} }
assoc_table assoc_table
} }
#[cfg(test)] #[cfg(test)]
mod tests { mod tests {
use super::*; use super::*;
@ -374,5 +374,3 @@ mod tests {
//maybe add tests relative to the matching optilization ? //maybe add tests relative to the matching optilization ?
} }