Garage v0.9 #473

Merged
lx merged 175 commits from next into main 2023-10-10 13:28:29 +00:00
2 changed files with 137 additions and 77 deletions
Showing only changes of commit bd842e1388 - Show all commits

View file

@ -206,7 +206,6 @@ impl Graph<FlowEdge>{
//There is no residual flow //There is no residual flow
break; break;
} }
//Now we run DFS respecting the level array //Now we run DFS respecting the level array
let mut next_nbd = vec![0; nb_vertices]; let mut next_nbd = vec![0; nb_vertices];
let mut lifo = VecDeque::new(); let mut lifo = VecDeque::new();
@ -220,14 +219,12 @@ impl Graph<FlowEdge>{
//The DFS reached the sink, we can add a //The DFS reached the sink, we can add a
//residual flow. //residual flow.
lifo.pop_back(); lifo.pop_back();
while !lifo.is_empty() { while let Some((id, _)) = lifo.pop_back() {
if let Some((id, _)) = lifo.pop_back() { let nbd = next_nbd[id];
let nbd = next_nbd[id]; self.graph[id][nbd].flow += f as i32;
self.graph[id][nbd].flow += f as i32; let id_rev = self.graph[id][nbd].dest;
let id_rev = self.graph[id][nbd].dest; let nbd_rev = self.graph[id][nbd].rev;
let nbd_rev = self.graph[id][nbd].rev; self.graph[id_rev][nbd_rev].flow -= f as i32;
self.graph[id_rev][nbd_rev].flow -= f as i32;
}
} }
lifo.push_back((idsource, flow_upper_bound)); lifo.push_back((idsource, flow_upper_bound));
continue; continue;
@ -243,10 +240,14 @@ impl Graph<FlowEdge>{
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, self.graph[id][nbd].cap - self.graph[id][nbd].flow as u32 ); let new_flow = min(f as i32, self.graph[id][nbd].cap as i32 - self.graph[id][nbd].flow) as u32;
if new_flow == 0 {
next_nbd[id] += 1;
continue;
}
if let (Some(lvldest), Some(lvlid)) = if let (Some(lvldest), Some(lvlid)) =
(level[self.graph[id][nbd].dest], level[id]){ (level[self.graph[id][nbd].dest], level[id]){
if lvldest <= lvlid || new_flow == 0 { if lvldest <= lvlid {
//We cannot send flow to nbd. //We cannot send flow to nbd.
next_nbd[id] += 1; next_nbd[id] += 1;
continue; continue;
@ -266,7 +267,6 @@ impl Graph<FlowEdge>{
// one needs to be present in the cost function. // one needs to be present in the cost function.
pub fn optimize_flow_with_cost(&mut self , cost: &CostFunction, path_length: usize ) pub fn optimize_flow_with_cost(&mut self , cost: &CostFunction, path_length: usize )
-> Result<(),String>{ -> Result<(),String>{
//We build the weighted graph g where we will look for negative cycle //We build the weighted graph g where we will look for negative cycle
let mut gf = self.build_cost_graph(cost)?; let mut gf = self.build_cost_graph(cost)?;
let mut cycles = gf.list_negative_cycles(path_length); let mut cycles = gf.list_negative_cycles(path_length);
@ -364,6 +364,7 @@ impl Graph<WeightedEdge>{
} }
} }
//If self.graph contains a negative cycle, then at this point the graph described //If self.graph contains a negative cycle, then at this point the graph described
//by prev (which is a directed 1-forest/functional graph) //by prev (which is a directed 1-forest/functional graph)
//must contain a cycle. We list the cycles of prev. //must contain a cycle. We list the cycles of prev.
@ -401,8 +402,9 @@ fn cycles_of_1_forest(forest: &[Option<usize>]) -> Vec<Vec<usize>> {
//We discovered an id that we explored at this iteration t. //We discovered an id that we explored at this iteration t.
//It means we are on a cycle //It means we are on a cycle
let mut cy = vec![id; 1]; let mut cy = vec![id; 1];
let id2 = id; let mut id2 = id;
while let Some(id2) = forest[id2] { while let Some(id_next) = forest[id2] {
id2 = id_next;
if id2 != id { if id2 != id {
cy.push(id2); cy.push(id2);
} }
@ -429,12 +431,5 @@ fn cycles_of_1_forest(forest: &[Option<usize>]) -> Vec<Vec<usize>> {
mod tests { mod tests {
use super::*; use super::*;
#[test]
fn test_flow() {
let left_vec = vec![3; 8];
let right_vec = vec![0, 4, 8, 4, 8];
//There are asserts in the function that computes the flow
}
//maybe add tests relative to the matching optilization ?
} }

View file

@ -3,6 +3,7 @@ use std::collections::HashMap;
use std::collections::HashSet; use std::collections::HashSet;
use hex::ToHex; use hex::ToHex;
use itertools::Itertools;
use serde::{Deserialize, Serialize}; use serde::{Deserialize, Serialize};
@ -185,7 +186,8 @@ impl ClusterLayout {
pub fn get_node_capacity(&self, uuid : &Uuid) -> Result<u32,String> { pub fn get_node_capacity(&self, uuid : &Uuid) -> Result<u32,String> {
match self.node_role(uuid) { match self.node_role(uuid) {
Some(NodeRole{capacity : Some(cap), zone: _, tags: _}) => return Ok(*cap), Some(NodeRole{capacity : Some(cap), zone: _, tags: _}) => return Ok(*cap),
_ => return Err("The Uuid does not correspond to a node present in the cluster or this node does not have a positive capacity.".to_string()) _ => return Err("The Uuid does not correspond to a node present in the \
cluster or this node does not have a positive capacity.".to_string())
} }
} }
@ -242,6 +244,47 @@ impl ClusterLayout {
} }
} }
//Check that every partition is associated to distinct nodes
let rf = self.replication_factor;
for p in 0..(1 << PARTITION_BITS) {
let nodes_of_p = self.ring_assignation_data[rf*p..rf*(p+1)].to_vec();
if nodes_of_p.iter().unique().count() != rf {
return false;
}
//Check that every partition is spread over at least zone_redundancy zones.
let zones_of_p = nodes_of_p.iter()
.map(|n| self.get_node_zone(&self.node_id_vec[*n as usize])
.expect("Zone not found."));
if zones_of_p.unique().count() < self.zone_redundancy {
return false;
}
}
//Check that the nodes capacities is consistent with the stored partitions
let mut node_usage = vec![0; MAX_NODE_NUMBER];
for n in self.ring_assignation_data.iter() {
node_usage[*n as usize] += 1;
}
for n in 0..MAX_NODE_NUMBER {
if node_usage[n] > 0 {
let uuid = self.node_id_vec[n];
if node_usage[n]*self.partition_size > self.get_node_capacity(&uuid)
.expect("Critical Error"){
return false;
}
}
}
//Check that the partition size stored is the one computed by the asignation
//algorithm.
let cl2 = self.clone();
let (_ , zone_to_id) = cl2.generate_zone_ids().expect("Critical Error");
let partition_size = cl2.compute_optimal_partition_size(&zone_to_id).expect("Critical Error");
if partition_size != self.partition_size {
return false;
}
true true
} }
@ -267,7 +310,7 @@ impl ClusterLayout {
self.zone_redundancy = redundancy; self.zone_redundancy = redundancy;
let mut msg = Message::new(); let mut msg = Message::new();
msg.push(format!("Computation of a new cluster layout where partitions are msg.push(format!("Computation of a new cluster layout where partitions are \
replicated {} times on at least {} distinct zones.", replication, redundancy)); replicated {} times on at least {} distinct zones.", replication, redundancy));
//We generate for once numerical ids for the zone, to use them as indices in the //We generate for once numerical ids for the zone, to use them as indices in the
@ -278,14 +321,17 @@ impl ClusterLayout {
self.useful_nodes().len(), id_to_zone.len())); self.useful_nodes().len(), id_to_zone.len()));
//We compute the optimal partition size //We compute the optimal partition size
//Capacities should be given in a unit so that partition size is at least 100.
//In this case, integer rounding plays a marginal role in the percentages of
//optimality.
let partition_size = self.compute_optimal_partition_size(&zone_to_id)?; let partition_size = self.compute_optimal_partition_size(&zone_to_id)?;
if old_assignation_opt != None { if old_assignation_opt != None {
msg.push(format!("Given the replication and redundancy constraint, the msg.push(format!("Given the replication and redundancy constraint, the \
optimal size of a partition is {}. In the previous layout, it used to optimal size of a partition is {}. In the previous layout, it used to \
be {}.", partition_size, self.partition_size)); be {}.", partition_size, self.partition_size));
} }
else { else {
msg.push(format!("Given the replication and redundancy constraints, the msg.push(format!("Given the replication and redundancy constraints, the \
optimal size of a partition is {}.", partition_size)); optimal size of a partition is {}.", partition_size));
} }
self.partition_size = partition_size; self.partition_size = partition_size;
@ -293,13 +339,13 @@ impl ClusterLayout {
//We compute a first flow/assignment that is heuristically close to the previous //We compute a first flow/assignment that is heuristically close to the previous
//assignment //assignment
let mut gflow = self.compute_candidate_assignment( &zone_to_id, &old_assignation_opt)?; let mut gflow = self.compute_candidate_assignment( &zone_to_id, &old_assignation_opt)?;
if let Some(assoc) = &old_assignation_opt { if let Some(assoc) = &old_assignation_opt {
//We minimize the distance to the previous assignment. //We minimize the distance to the previous assignment.
self.minimize_rebalance_load(&mut gflow, &zone_to_id, &assoc)?; self.minimize_rebalance_load(&mut gflow, &zone_to_id, &assoc)?;
} }
msg.append(&mut self.output_stat(&gflow, &old_assignation_opt, &zone_to_id,&id_to_zone)?); msg.append(&mut self.output_stat(&gflow, &old_assignation_opt, &zone_to_id,&id_to_zone)?);
msg.push("".to_string());
//We update the layout structure //We update the layout structure
self.update_ring_from_flow(id_to_zone.len() , &gflow)?; self.update_ring_from_flow(id_to_zone.len() , &gflow)?;
@ -321,7 +367,8 @@ impl ClusterLayout {
.map(|(k, _, _)| *k).collect(); .map(|(k, _, _)| *k).collect();
if new_non_gateway_nodes.len() > MAX_NODE_NUMBER { if new_non_gateway_nodes.len() > MAX_NODE_NUMBER {
return Err(format!("There are more than {} non-gateway nodes in the new layout. This is not allowed.", MAX_NODE_NUMBER).to_string()); return Err(format!("There are more than {} non-gateway nodes in the new \
layout. This is not allowed.", MAX_NODE_NUMBER).to_string());
} }
let mut new_gateway_nodes: Vec<Uuid> = self.roles.items().iter() let mut new_gateway_nodes: Vec<Uuid> = self.roles.items().iter()
@ -346,7 +393,8 @@ impl ClusterLayout {
return Ok(None); return Ok(None);
} }
if self.ring_assignation_data.len() != nb_partitions * self.replication_factor { if self.ring_assignation_data.len() != nb_partitions * self.replication_factor {
return Err("The old assignation does not have a size corresponding to the old replication factor or the number of partitions.".to_string()); return Err("The old assignation does not have a size corresponding to \
the old replication factor or the number of partitions.".to_string());
} }
//We build a translation table between the uuid and new ids //We build a translation table between the uuid and new ids
@ -384,7 +432,8 @@ impl ClusterLayout {
for uuid in self.node_id_vec.iter() { for uuid in self.node_id_vec.iter() {
if self.roles.get(uuid) == None { if self.roles.get(uuid) == None {
return Err("The uuid was not found in the node roles (this should not happen, it might be a critical error).".to_string()); return Err("The uuid was not found in the node roles (this should \
not happen, it might be a critical error).".to_string());
} }
match self.node_role(&uuid) { match self.node_role(&uuid) {
Some(r) => if !zone_to_id.contains_key(&r.zone) && r.capacity != None { Some(r) => if !zone_to_id.contains_key(&r.zone) && r.capacity != None {
@ -405,7 +454,8 @@ impl ClusterLayout {
let mut g = self.generate_flow_graph(1, zone_to_id, &empty_set)?; let mut g = self.generate_flow_graph(1, zone_to_id, &empty_set)?;
g.compute_maximal_flow()?; g.compute_maximal_flow()?;
if g.get_flow_value()? < (nb_partitions*self.replication_factor).try_into().unwrap() { if g.get_flow_value()? < (nb_partitions*self.replication_factor).try_into().unwrap() {
return Err("The storage capacity of he cluster is to small. It is impossible to store partitions of size 1.".to_string()); return Err("The storage capacity of he cluster is to small. It is \
impossible to store partitions of size 1.".to_string());
} }
let mut s_down = 1; let mut s_down = 1;
@ -525,7 +575,8 @@ impl ClusterLayout {
} }
if self.ring_assignation_data.len() != NB_PARTITIONS*self.replication_factor { if self.ring_assignation_data.len() != NB_PARTITIONS*self.replication_factor {
return Err("Critical Error : the association ring we produced does not have the right size.".to_string()); return Err("Critical Error : the association ring we produced does not \
have the right size.".to_string());
} }
return Ok(()); return Ok(());
} }
@ -546,9 +597,16 @@ impl ClusterLayout {
let percent_cap = 100.0*(used_cap as f32)/(total_cap as f32); let percent_cap = 100.0*(used_cap as f32)/(total_cap as f32);
msg.push(format!("Available capacity / Total cluster capacity: {} / {} ({:.1} %)", msg.push(format!("Available capacity / Total cluster capacity: {} / {} ({:.1} %)",
used_cap , total_cap , percent_cap )); used_cap , total_cap , percent_cap ));
msg.push(format!("If the percentage is to low, it might be that the replication/redundancy constraints force the use of nodes/zones with small storage capacities. msg.push(format!(""));
You might want to rebalance the storage capacities or relax the constraints. See the detailed statistics below and look for saturated nodes/zones.")); msg.push(format!("If the percentage is to low, it might be that the \
msg.push(format!("Recall that because of the replication, the actual available storage capacity is {} / {} = {}.", used_cap , self.replication_factor , used_cap/self.replication_factor as u32)); replication/redundancy constraints force the use of nodes/zones with small \
storage capacities. \
You might want to rebalance the storage capacities or relax the constraints. \
See the detailed statistics below and look for saturated nodes/zones."));
msg.push(format!("Recall that because of the replication, the actual available \
storage capacity is {} / {} = {}.",
used_cap , self.replication_factor ,
used_cap/self.replication_factor as u32));
//We define and fill in the following tables //We define and fill in the following tables
let storing_nodes = self.useful_nodes(); let storing_nodes = self.useful_nodes();
@ -563,6 +621,16 @@ impl ClusterLayout {
let pz_nodes = gflow.get_positive_flow_from(Vertex::PZ(p,z))?; let pz_nodes = gflow.get_positive_flow_from(Vertex::PZ(p,z))?;
if pz_nodes.len() > 0 { if pz_nodes.len() > 0 {
stored_partitions_zone[z] += 1; stored_partitions_zone[z] += 1;
if let Some(old_assoc) = old_assoc_opt {
let mut old_zones_of_p = Vec::<usize>::new();
for n in old_assoc[p].iter() {
old_zones_of_p.push(
zone_to_id[&self.get_node_zone(&self.node_id_vec[*n])?]);
}
if !old_zones_of_p.contains(&z) {
new_partitions_zone[z] += 1;
}
}
} }
for vert in pz_nodes.iter() { for vert in pz_nodes.iter() {
if let Vertex::N(n) = *vert { if let Vertex::N(n) = *vert {
@ -574,21 +642,17 @@ impl ClusterLayout {
} }
} }
} }
if let Some(old_assoc) = old_assoc_opt {
let mut old_zones_of_p = Vec::<usize>::new();
for n in old_assoc[p].iter() {
old_zones_of_p.push(
zone_to_id[&self.get_node_zone(&self.node_id_vec[*n])?]);
}
if !old_zones_of_p.contains(&z) {
new_partitions_zone[z] += 1;
}
}
} }
} }
if *old_assoc_opt == None {
new_partitions = stored_partitions.clone();
new_partitions_zone = stored_partitions_zone.clone();
}
//We display the statistics //We display the statistics
msg.push(format!(""));
if *old_assoc_opt != None { if *old_assoc_opt != None {
let total_new_partitions : usize = new_partitions.iter().sum(); let total_new_partitions : usize = new_partitions.iter().sum();
msg.push(format!("A total of {} new copies of partitions need to be \ msg.push(format!("A total of {} new copies of partitions need to be \
@ -608,16 +672,9 @@ impl ClusterLayout {
.map(|n| stored_partitions[*n]).sum(); .map(|n| stored_partitions[*n]).sum();
msg.push(format!("")); msg.push(format!(""));
if *old_assoc_opt != None { msg.push(format!("Zone {}: {} distinct partitions stored ({} new, \
msg.push(format!("Zone {}: {} distinct partitions stored ({} new, \
{} partition copies) ", id_to_zone[z], stored_partitions_zone[z], {} partition copies) ", id_to_zone[z], stored_partitions_zone[z],
new_partitions_zone[z], replicated_partitions)); new_partitions_zone[z], replicated_partitions));
}
else{
msg.push(format!("Zone {}: {} distinct partitions stored ({} partition \
copies) ",
id_to_zone[z], stored_partitions_zone[z], replicated_partitions));
}
let available_cap_z : u32 = self.partition_size*replicated_partitions as u32; let available_cap_z : u32 = self.partition_size*replicated_partitions as u32;
let mut total_cap_z = 0; let mut total_cap_z = 0;
@ -625,18 +682,17 @@ impl ClusterLayout {
total_cap_z += self.get_node_capacity(&self.node_id_vec[*n])?; total_cap_z += self.get_node_capacity(&self.node_id_vec[*n])?;
} }
let percent_cap_z = 100.0*(available_cap_z as f32)/(total_cap_z as f32); let percent_cap_z = 100.0*(available_cap_z as f32)/(total_cap_z as f32);
msg.push(format!(" Available capacity / Total capacity: {}/{} ({:.1}%).", msg.push(format!(" Available capacity / Total capacity: {}/{} ({:.1}%).",
available_cap_z, total_cap_z, percent_cap_z)); available_cap_z, total_cap_z, percent_cap_z));
msg.push(format!(""));
for n in nodes_of_z.iter() { for n in nodes_of_z.iter() {
let available_cap_n = stored_partitions[*n] as u32 *self.partition_size; let available_cap_n = stored_partitions[*n] as u32 *self.partition_size;
let total_cap_n =self.get_node_capacity(&self.node_id_vec[*n])?; let total_cap_n =self.get_node_capacity(&self.node_id_vec[*n])?;
let tags_n = (self.node_role(&self.node_id_vec[*n]) let tags_n = (self.node_role(&self.node_id_vec[*n])
.ok_or("Node not found."))?.tags_string(); .ok_or("Node not found."))?.tags_string();
msg.push(format!(" Node {}: {} partitions ({} new) ; \ msg.push(format!(" Node {}: {} partitions ({} new) ; \
available/total capacity: {} / {} ({:.1}%) ; tags:{}", available/total capacity: {} / {} ({:.1}%) ; tags:{}",
&self.node_id_vec[*n].to_vec().encode_hex::<String>(), &self.node_id_vec[*n].to_vec()[0..2].to_vec().encode_hex::<String>(),
stored_partitions[*n], stored_partitions[*n],
new_partitions[*n], available_cap_n, total_cap_n, new_partitions[*n], available_cap_n, total_cap_n,
(available_cap_n as f32)/(total_cap_n as f32)*100.0 , (available_cap_n as f32)/(total_cap_n as f32)*100.0 ,
@ -654,16 +710,14 @@ impl ClusterLayout {
#[cfg(test)] #[cfg(test)]
mod tests { mod tests {
use super::*; use super::*;
use itertools::Itertools; use std::io::*;
// 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!(cl.ring_assignation_data.len() == nb_partitions * cl.replication_factor); assert!(cl.ring_assignation_data.len() == nb_partitions * cl.replication_factor);
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 {
@ -743,6 +797,13 @@ mod tests {
} }
} }
} }
*/
fn show_msg(msg : &Message) {
for s in msg.iter(){
println!("{}",s);
}
}
fn update_layout( fn update_layout(
cl: &mut ClusterLayout, cl: &mut ClusterLayout,
@ -769,7 +830,8 @@ mod tests {
#[test] #[test]
fn test_assignation() { fn test_assignation() {
let mut node_id_vec = vec![1, 2, 3]; std::io::stdout().flush().ok().expect("Could not flush stdout");
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"] let mut node_zone_vec = vec!["A", "B", "C"]
.into_iter() .into_iter()
@ -782,14 +844,16 @@ mod tests {
roles: LwwMap::new(), roles: LwwMap::new(),
replication_factor: 3, replication_factor: 3,
zone_redundancy: 1,
partition_size: 0,
ring_assignation_data: vec![], ring_assignation_data: vec![],
version: 0, version: 0,
staging: LwwMap::new(), staging: LwwMap::new(),
staging_hash: sha256sum(&[1; 32]), staging_hash: blake2sum(&rmp_to_vec_all_named(&LwwMap::<Uuid, NodeRoleV>::new()).unwrap()[..]),
}; };
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(); show_msg(&cl.calculate_partition_assignation(3,3).unwrap());
check_assignation(&cl); assert!(cl.check());
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];
@ -798,17 +862,18 @@ mod tests {
.map(|x| x.to_string()) .map(|x| x.to_string())
.collect(); .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(); show_msg(&cl.calculate_partition_assignation(3,3).unwrap());
check_assignation(&cl); assert!(cl.check());
node_capacity_vec = vec![4000, 1000, 2000, 7000, 1000, 1000, 2000, 10000, 2000]; node_capacity_vec = vec![4000, 1000, 2000, 7000, 1000, 1000, 2000, 10000, 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(); show_msg(&cl.calculate_partition_assignation(3,3).unwrap());
check_assignation(&cl); assert!(cl.check());
node_capacity_vec = vec![4000, 4000, 2000, 7000, 1000, 9000, 2000, 10, 2000]; node_capacity_vec = vec![4000000, 4000000, 2000000, 7000000, 1000000, 9000000, 2000000, 10000, 2000000];
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(); show_msg(&cl.calculate_partition_assignation(3,1).unwrap());
check_assignation(&cl); assert!(cl.check());
} }
} }