netmap: sort buckets by HRW value

If weights for some buckets are equal, result depends on the
initial positions of buckets in slice. Thus we sort buckets by value
using hash of the first node as a have for the whole bucket.
This is ok, because buckets are already sorted by HRW.

Signed-off-by: Evgenii Stratonikov <evgeniy@nspcc.ru>
This commit is contained in:
Evgenii Stratonikov 2021-12-24 14:31:34 +03:00 committed by Alex Vanin
parent f7582399ed
commit 3edaf9ecb6
3 changed files with 192 additions and 2 deletions

View file

@ -96,6 +96,16 @@ func (n Node) Hash() uint64 {
return n.ID return n.ID
} }
// Hash is a function from hrw.Hasher interface. It is implemented
// to support weighted hrw sorting of buckets. Each bucket is already sorted by hrw,
// thus giving us needed "randomness".
func (n Nodes) Hash() uint64 {
if len(n) > 0 {
return n[0].Hash()
}
return 0
}
// NodesFromInfo converts slice of NodeInfo to a generic node slice. // NodesFromInfo converts slice of NodeInfo to a generic node slice.
func NodesFromInfo(infos []NodeInfo) Nodes { func NodesFromInfo(infos []NodeInfo) Nodes {
nodes := make(Nodes, len(infos)) nodes := make(Nodes, len(infos))

View file

@ -58,8 +58,11 @@ func (c *Context) getSelection(p *PlacementPolicy, s *Selector) ([]Nodes, error)
return nil, fmt.Errorf("%w: '%s'", ErrNotEnoughNodes, s.Name()) return nil, fmt.Errorf("%w: '%s'", ErrNotEnoughNodes, s.Name())
} }
// We need deterministic output in case there is no pivot.
// If pivot is set, buckets are sorted by HRW.
// However, because initial order influences HRW order for buckets with equal weights,
// we also need to have deterministic input to HRW sorting routine.
if len(c.pivot) == 0 { if len(c.pivot) == 0 {
// Deterministic order in case of zero seed.
if s.Attribute() == "" { if s.Attribute() == "" {
sort.Slice(buckets, func(i, j int) bool { sort.Slice(buckets, func(i, j int) bool {
return buckets[i].nodes[0].ID < buckets[j].nodes[0].ID return buckets[i].nodes[0].ID < buckets[j].nodes[0].ID
@ -98,7 +101,7 @@ func (c *Context) getSelection(p *PlacementPolicy, s *Selector) ([]Nodes, error)
weights[i] = GetBucketWeight(nodes[i], c.aggregator(), c.weightFunc) weights[i] = GetBucketWeight(nodes[i], c.aggregator(), c.weightFunc)
} }
hrw.SortSliceByWeightIndex(nodes, weights, c.pivotHash) hrw.SortSliceByWeightValue(nodes, weights, c.pivotHash)
} }
if s.Attribute() == "" { if s.Attribute() == "" {

View file

@ -3,13 +3,190 @@ package netmap
import ( import (
"errors" "errors"
"fmt" "fmt"
"math/rand"
"sort"
"strconv"
"testing" "testing"
"github.com/nspcc-dev/hrw"
"github.com/nspcc-dev/neofs-api-go/v2/netmap" "github.com/nspcc-dev/neofs-api-go/v2/netmap"
testv2 "github.com/nspcc-dev/neofs-api-go/v2/netmap/test" testv2 "github.com/nspcc-dev/neofs-api-go/v2/netmap/test"
"github.com/stretchr/testify/require" "github.com/stretchr/testify/require"
) )
func BenchmarkHRWSort(b *testing.B) {
const netmapSize = 1000
nodes := make([]Nodes, netmapSize)
weights := make([]float64, netmapSize)
for i := range nodes {
nodes[i] = Nodes{{
ID: rand.Uint64(),
Index: i,
Capacity: 100,
Price: 1,
AttrMap: nil,
}}
weights[i] = float64(rand.Uint32()%10) / 10.0
}
pivot := rand.Uint64()
b.Run("sort by index, no weight", func(b *testing.B) {
realNodes := make([]Nodes, netmapSize)
b.ResetTimer()
for i := 0; i < b.N; i++ {
b.StopTimer()
copy(realNodes, nodes)
b.StartTimer()
hrw.SortSliceByIndex(realNodes, pivot)
}
})
b.Run("sort by value, no weight", func(b *testing.B) {
realNodes := make([]Nodes, netmapSize)
b.ResetTimer()
for i := 0; i < b.N; i++ {
b.StopTimer()
copy(realNodes, nodes)
b.StartTimer()
hrw.SortSliceByValue(realNodes, pivot)
}
})
b.Run("only sort by index", func(b *testing.B) {
realNodes := make([]Nodes, netmapSize)
b.ResetTimer()
for i := 0; i < b.N; i++ {
b.StopTimer()
copy(realNodes, nodes)
b.StartTimer()
hrw.SortSliceByWeightIndex(realNodes, weights, pivot)
}
})
b.Run("sort by value", func(b *testing.B) {
realNodes := make([]Nodes, netmapSize)
b.ResetTimer()
for i := 0; i < b.N; i++ {
b.StopTimer()
copy(realNodes, nodes)
b.StartTimer()
hrw.SortSliceByWeightValue(realNodes, weights, pivot)
}
})
b.Run("sort by ID, then by index (deterministic)", func(b *testing.B) {
realNodes := make([]Nodes, netmapSize)
b.ResetTimer()
for i := 0; i < b.N; i++ {
b.StopTimer()
copy(realNodes, nodes)
b.StartTimer()
sort.Slice(nodes, func(i, j int) bool {
return nodes[i][0].ID < nodes[j][0].ID
})
hrw.SortSliceByWeightIndex(realNodes, weights, pivot)
}
})
}
func BenchmarkPolicyHRWType(b *testing.B) {
const netmapSize = 100
p := newPlacementPolicy(1,
[]*Replica{
newReplica(1, "loc1"),
newReplica(1, "loc2")},
[]*Selector{
newSelector("loc1", "Location", ClauseSame, 1, "loc1"),
newSelector("loc2", "Location", ClauseSame, 1, "loc2")},
[]*Filter{
newFilter("loc1", "Location", "Shanghai", OpEQ),
newFilter("loc2", "Location", "Shanghai", OpNE),
})
nodes := make([]NodeInfo, netmapSize)
for i := range nodes {
var loc string
switch i % 20 {
case 0:
loc = "Shanghai"
default:
loc = strconv.Itoa(i % 20)
}
// Having the same price and capacity ensures equal weights for all nodes.
// This way placement is more dependent on the initial order.
nodes[i] = nodeInfoFromAttributes("Location", loc, "Price", "1", "Capacity", "10")
pub := make([]byte, 33)
pub[0] = byte(i)
nodes[i].SetPublicKey(pub)
}
nm, err := NewNetmap(NodesFromInfo(nodes))
require.NoError(b, err)
b.ResetTimer()
for i := 0; i < b.N; i++ {
_, err := nm.GetContainerNodes(p, []byte{1})
if err != nil {
b.Fatal()
}
}
}
func TestPlacementPolicy_DeterministicOrder(t *testing.T) {
const netmapSize = 100
p := newPlacementPolicy(1,
[]*Replica{
newReplica(1, "loc1"),
newReplica(1, "loc2")},
[]*Selector{
newSelector("loc1", "Location", ClauseSame, 1, "loc1"),
newSelector("loc2", "Location", ClauseSame, 1, "loc2")},
[]*Filter{
newFilter("loc1", "Location", "Shanghai", OpEQ),
newFilter("loc2", "Location", "Shanghai", OpNE),
})
nodes := make([]NodeInfo, netmapSize)
for i := range nodes {
var loc string
switch i % 20 {
case 0:
loc = "Shanghai"
default:
loc = strconv.Itoa(i % 20)
}
// Having the same price and capacity ensures equal weights for all nodes.
// This way placement is more dependent on the initial order.
nodes[i] = nodeInfoFromAttributes("Location", loc, "Price", "1", "Capacity", "10")
pub := make([]byte, 33)
pub[0] = byte(i)
nodes[i].SetPublicKey(pub)
}
nm, err := NewNetmap(NodesFromInfo(nodes))
require.NoError(t, err)
getIndices := func(t *testing.T) (int, int) {
v, err := nm.GetContainerNodes(p, []byte{1})
require.NoError(t, err)
ns := v.Flatten()
require.Equal(t, 2, len(ns))
return ns[0].Index, ns[1].Index
}
a, b := getIndices(t)
for i := 0; i < 10; i++ {
x, y := getIndices(t)
require.Equal(t, a, x)
require.Equal(t, b, y)
}
}
func TestPlacementPolicy_UnspecifiedClause(t *testing.T) { func TestPlacementPolicy_UnspecifiedClause(t *testing.T) {
p := newPlacementPolicy(1, p := newPlacementPolicy(1,
[]*Replica{newReplica(1, "X")}, []*Replica{newReplica(1, "X")},