1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
|
// Copyright Verizon Media. Licensed under the terms of the Apache 2.0 license. See LICENSE in the project root.
package com.yahoo.vespa.hosted.provision.autoscale;
import com.yahoo.config.provision.ClusterResources;
import com.yahoo.config.provision.NodeResources;
import com.yahoo.vespa.hosted.provision.NodeRepository;
import java.util.Optional;
/**
* A searcher of the space of possible allocation
*
* @author bratseth
*/
public class AllocationOptimizer {
// The min and max nodes to consider when not using application supplied limits
private static final int minimumNodes = 2; // Since this number includes redundancy it cannot be lower than 2
private static final int maximumNodes = 150;
// When a query is issued on a node the cost is the sum of a fixed cost component and a cost component
// proportional to document count. We must account for this when comparing configurations with more or fewer nodes.
// TODO: Measure this, and only take it into account with queries
private static final double fixedCpuCostFraction = 0.1;
private final NodeRepository nodeRepository;
public AllocationOptimizer(NodeRepository nodeRepository) {
this.nodeRepository = nodeRepository;
}
/**
* Searches the space of possible allocations given a target
* and (optionally) cluster limits and returns the best alternative.
*
* @return the best allocation, if there are any possible legal allocations, fulfilling the target
* fully or partially, within the limits
*/
public Optional<AllocatableClusterResources> findBestAllocation(ResourceTarget target,
AllocatableClusterResources current,
Limits limits,
boolean exclusive) {
if (limits.isEmpty())
limits = Limits.of(new ClusterResources(minimumNodes, 1, NodeResources.unspecified()),
new ClusterResources(maximumNodes, maximumNodes, NodeResources.unspecified()));
Optional<AllocatableClusterResources> bestAllocation = Optional.empty();
for (int groups = limits.min().groups(); groups <= limits.max().groups(); groups++) {
for (int nodes = limits.min().nodes(); nodes <= limits.max().nodes(); nodes++) {
if (nodes % groups != 0) continue;
int groupSize = nodes / groups;
// Adjust for redundancy: Node in group if groups = 1, an extra group if multiple groups
// TODO: Make the best choice based on size and redundancy setting instead
int nodesAdjustedForRedundancy = target.adjustForRedundancy() ? (groups == 1 ? nodes - 1 : nodes - groupSize) : nodes;
int groupsAdjustedForRedundancy = target.adjustForRedundancy() ? (groups == 1 ? 1 : groups - 1) : groups;
ClusterResources next = new ClusterResources(nodes,
groups,
nodeResourcesWith(nodesAdjustedForRedundancy, groupsAdjustedForRedundancy, limits, current, target));
var allocatableResources = AllocatableClusterResources.from(next, exclusive, current.clusterType(), limits, nodeRepository);
if (allocatableResources.isEmpty()) continue;
if (bestAllocation.isEmpty() || allocatableResources.get().preferableTo(bestAllocation.get()))
bestAllocation = allocatableResources;
}
}
return bestAllocation;
}
/**
* For the observed load this instance is initialized with, returns the resources needed per node to be at
* ideal load given a target node count
*/
private NodeResources nodeResourcesWith(int nodes, int groups, Limits limits, AllocatableClusterResources current, ResourceTarget target) {
// Cpu: Scales with cluster size (TODO: Only reads, writes scales with group size)
// Memory and disk: Scales with group size
double cpu, memory, disk;
int groupSize = nodes / groups;
if (current.clusterType().isContent()) { // load scales with node share of content
// The fixed cost portion of cpu does not scale with changes to the node count
// TODO: Only for the portion of cpu consumed by queries
double cpuPerGroup = fixedCpuCostFraction * target.nodeCpu() +
(1 - fixedCpuCostFraction) * target.nodeCpu() * current.groupSize() / groupSize;
cpu = cpuPerGroup * current.groups() / groups;
memory = target.nodeMemory() * current.groupSize() / groupSize;
disk = target.nodeDisk() * current.groupSize() / groupSize;
}
else {
cpu = target.nodeCpu() * current.nodes() / nodes;
memory = target.nodeMemory();
disk = target.nodeDisk();
}
// Combine the scaled resource values computed here
// with the currently configured non-scaled values, given in the limits, if any
NodeResources nonScaled = limits.isEmpty() || limits.min().nodeResources().isUnspecified()
? current.toAdvertisedClusterResources().nodeResources()
: limits.min().nodeResources(); // min=max for non-scaled
return nonScaled.withVcpu(cpu).withMemoryGb(memory).withDiskGb(disk);
}
}
|