Class: Ignis::Collective::Algorithms::TopologyRouter

Inherits:
Object
  • Object
show all
Defined in:
lib/nvruby/collective/algorithms/topology_router.rb

Overview

Topology-Aware Routing Optimizer

Optimizes collective operation routing based on PCIe/NVLink topology. Groups GPUs by PCIe switch to minimize cross-switch traffic.

Key optimizations:

  1. Intra-switch communication first (lower latency)

  2. Ring order follows physical topology

  3. Tree structure respects switch boundaries

Defined Under Namespace

Classes: SwitchGroup

Instance Attribute Summary collapse

Instance Method Summary collapse

Constructor Details

#initialize(gpu_ids:, topology: nil) ⇒ TopologyRouter

Returns a new instance of TopologyRouter.

Parameters:

  • gpu_ids (Array<Integer>) —

    GPU device IDs

  • topology (Topology::Detector, nil) (defaults to: nil) —

    Optional topology detector



30
31
32
33
34
35
36
 
# File 'lib/nvruby/collective/algorithms/topology_router.rb', line 30

def initialize(gpu_ids:, topology: nil)
  @gpu_ids = gpu_ids.dup.freeze
  @topology = topology || Topology::Detector.new
  @switch_groups = []
  @routing_table = {}
  detect_switch_groups!
end

Instance Attribute Details

#gpu_ids ⇒ Array<Integer> (readonly)

Returns GPU IDs.

Returns:

  • (Array<Integer>) —

    GPU IDs



23
24
25
 
# File 'lib/nvruby/collective/algorithms/topology_router.rb', line 23

def gpu_ids
  @gpu_ids
end

#switch_groups ⇒ Array<SwitchGroup> (readonly)

Returns GPU groups by switch.

Returns:



26
27
28
 
# File 'lib/nvruby/collective/algorithms/topology_router.rb', line 26

def switch_groups
  @switch_groups
end

#topology ⇒ Topology::Detector (readonly)

Returns Topology detector.

Returns:



20
21
22
 
# File 'lib/nvruby/collective/algorithms/topology_router.rb', line 20

def topology
  @topology
end

Instance Method Details

#get_path(src_rank, dst_rank) ⇒ Array<Integer>

Get routing path between two GPUs

Parameters:

  • src_rank (Integer) —

    Source rank

  • dst_rank (Integer) —

    Destination rank

Returns:

  • (Array<Integer>) —

    Path as list of rank hops



153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
 
# File 'lib/nvruby/collective/algorithms/topology_router.rb', line 153

def get_path(src_rank, dst_rank)
  return [src_rank] if src_rank == dst_rank

  src_gpu = @gpu_ids[src_rank]
  dst_gpu = @gpu_ids[dst_rank]

  src_group = find_group(src_gpu)
  dst_group = find_group(dst_gpu)

  if src_group == dst_group
    # Direct path within same switch
    [src_rank, dst_rank]
  else
    # Need to cross switch boundary - use group representative
    # Find best intermediate hop
    intermediate = pick_best_connector_rank(src_group, dst_group)
    
    if intermediate == src_rank || intermediate == dst_rank
      [src_rank, dst_rank]
    else
      [src_rank, intermediate, dst_rank]
    end
  end
end

#optimal_ring_order ⇒ Array<Integer>

Get optimized ring order based on topology Groups GPUs by PCIe switch, then chains switches

Returns:

  • (Array<Integer>) —

    Optimal ring order



42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
 
# File 'lib/nvruby/collective/algorithms/topology_router.rb', line 42

def optimal_ring_order
  return @gpu_ids if @switch_groups.size <= 1

  # Build ring by chaining switch groups
  # Within each group, order by P2P performance
  ring = []

  # Sort groups by average bandwidth (best first)
  sorted_groups = @switch_groups.sort_by { |g| -g.bandwidth_gbps }

  sorted_groups.each do |group|
    # Order GPUs within group by P2P performance
    ordered = order_within_group(group.gpu_ids)
    ring.concat(ordered)
  end

  ring
end

#optimal_tree_structure(root: 0) ⇒ Hash

Get optimized tree structure that respects switch boundaries

Parameters:

  • root (Integer) (defaults to: 0) —

    Root rank

Returns:

  • (Hash) —

    Tree structure with :parent, :children for each rank



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
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
 
# File 'lib/nvruby/collective/algorithms/topology_router.rb', line 65

def optimal_tree_structure(root: 0)
  tree = {}
  n = @gpu_ids.size

  return { 0 => { parent: nil, children: [] } } if n == 1

  # Build hierarchical tree:
  # - Level 0: Root
  # - Level 1: One representative from each switch group
  # - Level 2+: Other GPUs within each switch group

  # Find which group the root belongs to
  root_group_idx = @switch_groups.find_index { |g| g.gpu_ids.include?(@gpu_ids[root]) }
  root_group = @switch_groups[root_group_idx]

  # Root has no parent
  tree[root] = { parent: nil, children: [] }

  # Connect switch group representatives to root
  group_reps = []
  @switch_groups.each_with_index do |group, idx|
    next if idx == root_group_idx

    # Pick highest-bandwidth GPU in group as representative
    rep_gpu = pick_best_connector(root_group, group)
    rep_rank = @gpu_ids.index(rep_gpu)
    
    tree[root][:children] << rep_rank
    tree[rep_rank] = { parent: root, children: [] }
    group_reps << { rank: rep_rank, group: group }
  end

  # Connect other GPUs in root's group directly to root
  root_group.gpu_ids.each do |gpu|
    rank = @gpu_ids.index(gpu)
    next if rank == root
    
    tree[root][:children] << rank
    tree[rank] = { parent: root, children: [] }
  end

  # Connect remaining GPUs in each group to their representative
  group_reps.each do |rep_info|
    rep_rank = rep_info[:rank]
    rep_gpu = @gpu_ids[rep_rank]
    
    rep_info[:group].gpu_ids.each do |gpu|
      next if gpu == rep_gpu
      
      rank = @gpu_ids.index(gpu)
      tree[rep_rank][:children] << rank
      tree[rank] = { parent: rep_rank, children: [] }
    end
  end

  tree
end

#suggest_algorithm(message_size) ⇒ Symbol

Suggest optimal algorithm based on message size and topology

Parameters:

  • message_size (Integer) —

    Total message size in bytes

Returns:

  • (Symbol) —

    :ring, :tree, or :double_tree



127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
 
# File 'lib/nvruby/collective/algorithms/topology_router.rb', line 127

def suggest_algorithm(message_size)
  # Ring is bandwidth-optimal for large messages
  # Tree is latency-optimal for small messages
  # Double tree for non-power-of-2 with medium messages

  n = @gpu_ids.size
  is_power_of_2 = n > 0 && (n & (n - 1)).zero?

  if message_size < 1024  # < 1KB
    :tree
  elsif message_size > 1_048_576  # > 1MB
    :ring
  elsif !is_power_of_2
    :double_tree
  else
    # Medium messages: prefer ring if good interconnect
    avg_bandwidth = @switch_groups.sum { |g| g.bandwidth_gbps } / @switch_groups.size.to_f
    avg_bandwidth > 50 ? :ring : :tree
  end
end