Find multi-node NVLink domains, place across them, and show them as racks

A GB200 NVL72 rack or an NVLink Switch system is a scale-up island: its
GPUs reach each other over NVLink whatever node they are in. Each GPU
now reports its fabric from nvidia-smi -q - state, status, ClusterUUID,
CliqueId, bandwidth - at registration and once a minute after, since a
fabric manager restarting drops every card out at once; and each node
reports nvidia-imex and its channels, without which the domain's GPUs
cannot share memory across nodes. A single machine's NVSwitch, which
reports a cluster of zeros, is no domain.

A replica spanning nodes is placed inside one domain first, on nodes
whose every card is joined and whose IMEX is up, and its shards are
told NCCL_MNNVL_ENABLE=1; InfiniBand is the fallback.

/admin/nvlink draws each domain as a rack - the NVLink Switch band on
top, light running down a backplane, trays of GPU chips that glow when
joined, pulse red when not, turn amber on degraded bandwidth - and
lists what keeps a tray out. A node's page shows its GPU-by-GPU NVLink
matrix, each card's NUMA node and nearest adapter, and its domain and
IMEX. doctor checks joining, bandwidth, IMEX and its channel; metrics
llmgw_gpu_nvlink_joined and llmgw_imex_up; /api/v1/nvlink.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
This commit is contained in:
Ilker ManapandClaude Opus 5.5 committed 2026-09-29 09:07:59 +02:00
1 parent 5bbc619a31
commit 20479a2d5f
31 files changed
+1535 -131

No files matched your search

+46
View File
@@ -16,11 +16,13 @@ import (
"strings"
"time"
"gitea.mynodes.xyz/ilker/llmgw/internal/agent/discovery"
"gitea.mynodes.xyz/ilker/llmgw/internal/bench"
"gitea.mynodes.xyz/ilker/llmgw/internal/control"
"gitea.mynodes.xyz/ilker/llmgw/internal/control/catalog"
"gitea.mynodes.xyz/ilker/llmgw/internal/control/fabrics"
"gitea.mynodes.xyz/ilker/llmgw/internal/control/inventory"
"gitea.mynodes.xyz/ilker/llmgw/internal/control/nvlink"
"gitea.mynodes.xyz/ilker/llmgw/internal/control/reconciler"
"gitea.mynodes.xyz/ilker/llmgw/internal/control/topology"
"gitea.mynodes.xyz/ilker/llmgw/internal/store"
@@ -53,6 +55,9 @@ func (a *API) Register(mux *http.ServeMux) {
mux.HandleFunc("GET /api/v1/fabrics", func(w http.ResponseWriter, _ *http.Request) {
writeJSON(w, http.StatusOK, a.Fabrics())
})
mux.HandleFunc("GET /api/v1/nvlink", func(w http.ResponseWriter, _ *http.Request) {
writeJSON(w, http.StatusOK, a.NVLink())
})
mux.HandleFunc("GET /api/v1/fabrics/{id}/topology", func(w http.ResponseWriter, r *http.Request) {
t, ok := a.Topology(r.PathValue("id"))
if !ok {
@@ -134,6 +139,27 @@ type NodeView struct {
Load *LoadView `json:"load,omitempty"`
// IBPorts are the node's InfiniBand and RoCE ports at the last heartbeat.
IBPorts []IBPortView `json:"ib_ports,omitempty"`
// GPUs are the cards, with how they are wired: NVLink peers, the
// adapter nearest each, and the NVLink domain each has joined.
GPUs []GPUView `json:"gpus,omitempty"`
// NVLinkDomain is the multi-node domain every card has joined, if any.
NVLinkDomain string `json:"nvlink_domain,omitempty"`
IMEX string `json:"imex,omitempty"`
IMEXChannels int `json:"imex_channels,omitempty"`
}
// GPUView is one card as a node's page shows it.
type GPUView struct {
Index int `json:"index"`
Name string `json:"name"`
NUMA int `json:"numa"`
NVLinkPeers []int `json:"nvlink_peers,omitempty"`
// NearestNIC is the adapter closest to the card, and how close.
NearestNIC string `json:"nearest_nic,omitempty"`
NICPath string `json:"nic_path,omitempty"`
FabricState string `json:"fabric_state,omitempty"`
Domain string `json:"domain,omitempty"`
Bandwidth string `json:"bandwidth,omitempty"`
}
// IBPortView is one port as a page shows it.
@@ -344,6 +370,20 @@ func (a *API) Nodes() []NodeView {
v.LastSeenAgo = time.Since(n.LastSeen).Round(time.Second).String()
}
v.Load = loadView(n.Load, time.Now())
v.NVLinkDomain, v.IMEX, v.IMEXChannels = n.NVLinkDomain(), n.IMEX.State, n.IMEX.Channels
for _, g := range n.GPUs {
gv := GPUView{Index: g.Index, Name: g.Name, NUMA: g.NUMANode, NVLinkPeers: g.NVLinkPeers}
for hca, rel := range g.NICPaths {
if gv.NearestNIC == "" || discovery.PathRank(rel) < discovery.PathRank(gv.NICPath) ||
(discovery.PathRank(rel) == discovery.PathRank(gv.NICPath) && hca < gv.NearestNIC) {
gv.NearestNIC, gv.NICPath = hca, rel
}
}
if f := g.Fabric; f != nil {
gv.FabricState, gv.Domain, gv.Bandwidth = f.State, f.Domain(), f.Bandwidth
}
v.GPUs = append(v.GPUs, gv)
}
fabricOf := map[string]string{}
for _, f := range n.Fabrics {
for _, h := range f.HCAs {
@@ -959,3 +999,9 @@ func (a *API) Topology(id string) (TopologyView, bool) {
}
return t, true
}
// NVLink is the cluster's multi-node NVLink domains, and what keeps any of
// their nodes out.
func (a *API) NVLink() nvlink.Report {
return nvlink.Analyze(a.ctl.Inventory().Nodes())
}
+18
View File
@@ -61,6 +61,8 @@ func NodeFromRegister(r *llmgwv1.Register) inventory.Node {
CUDAVersion: g.GetCudaVersion(),
NVLinkPeers: intSlice(g.GetNvlinkPeers()),
NICPaths: maps.Clone(g.GetNicPaths()),
Fabric: fabricFrom(g.GetFabricState(), g.GetFabricStatus(), g.GetClusterUuid(),
g.GetCliqueId(), g.GetFabricBandwidth()),
Health: discovery.GPUHealth{
ECCErrors: int(g.GetEccErrors()),
TemperatureC: int(g.GetTemperatureC()),
@@ -68,6 +70,7 @@ func NodeFromRegister(r *llmgwv1.Register) inventory.Node {
},
})
}
n.IMEX = discovery.IMEX{State: r.GetImex().GetState(), Channels: int(r.GetImex().GetChannels())}
for _, f := range r.GetFabrics() {
n.Fabrics = append(n.Fabrics, inventory.FabricAttachment{
Fabric: f.GetFabric(),
@@ -119,7 +122,13 @@ func RegisterFromInventory(nodeID, hostname, mgmtAddr, agentVersion string,
TemperatureC: int32(g.Health.TemperatureC),
EccErrors: int32(g.Health.ECCErrors),
})
if f := g.Fabric; f != nil {
last := r.Gpus[len(r.Gpus)-1]
last.FabricState, last.FabricStatus, last.ClusterUuid = f.State, f.Status, f.ClusterUUID
last.CliqueId, last.FabricBandwidth = f.CliqueID, f.Bandwidth
}
}
r.Imex = &llmgwv1.IMEX{State: inv.IMEX.State, Channels: int32(inv.IMEX.Channels)}
for _, f := range fabrics {
r.Fabrics = append(r.Fabrics, &llmgwv1.FabricAttachment{
Fabric: f.Fabric, Hcas: f.HCAs, RateGbps: int32(f.RateGbps),
@@ -266,3 +275,12 @@ func int32Slice(v []int) []int32 {
}
return out
}
// fabricFrom is a GPU's NVLink fabric as the wire carries it, nil when the
// driver reported none.
func fabricFrom(state, status, cluster, clique, bandwidth string) *discovery.GPUFabric {
if state == "" {
return nil
}
return &discovery.GPUFabric{State: state, Status: status, ClusterUUID: cluster, CliqueID: clique, Bandwidth: bandwidth}
}
@@ -242,3 +242,25 @@ func TestAPortGoingDownTakesItsFabricAway(t *testing.T) {
t.Error("a heartbeat without a fabric change emptied the fabrics")
}
}
func TestAGPUFallingOutOfItsDomainTakesTheNodeOut(t *testing.T) {
// A fabric manager restarting drops cards out of the domain; the node has
// to stop being offered work that spans it until they are back.
s := newTestService()
fab := &discovery.GPUFabric{State: "Completed", ClusterUUID: "76b1ad76-7c24", CliqueID: "1"}
nodeWith(s, inventory.Node{ID: "t1", State: inventory.Ready, GPUs: []discovery.GPU{
{Index: 0, Fabric: fab}, {Index: 1, Fabric: fab}}})
if n, _ := s.inv.Get("t1"); n.NVLinkDomain() == "" {
t.Fatal("the node should start in its domain")
}
s.applyHeartbeat("t1", &llmgwv1.Heartbeat{NodeId: "t1",
GpuFabric: []*llmgwv1.GPUFabricState{{Index: 1, State: "In Progress"}},
Imex: &llmgwv1.IMEX{State: "active", Channels: 1}})
n, _ := s.inv.Get("t1")
if n.NVLinkDomain() != "" {
t.Errorf("a node with a card out of the domain is still in it")
}
if n.IMEX.State != "active" || n.IMEX.Channels != 1 {
t.Errorf("imex = %+v", n.IMEX)
}
}
+19 -1
View File
@@ -322,7 +322,7 @@ func (s *Service) applyHeartbeat(nodeID string, hb *llmgwv1.Heartbeat) {
if len(used) == 0 && len(problems) == 0 && len(storage) == 0 &&
!hb.GetWeightsChanged() && memMB == 0 && !hb.GetEnginesChanged() &&
!hb.GetProvisionChanged() && hb.GetLoad() == nil && len(hb.GetIbPorts()) == 0 &&
!hb.GetFabricsChanged() {
!hb.GetFabricsChanged() && len(hb.GetGpuFabric()) == 0 && hb.GetImex() == nil {
return
}
@@ -457,6 +457,24 @@ func (s *Service) applyHeartbeat(nodeID string, hb *llmgwv1.Heartbeat) {
if len(hb.GetIbPorts()) > 0 {
n.IBAt = time.Now()
}
// NVLink domain membership, which a fabric manager restarting takes
// away from every card at once.
if fab := hb.GetGpuFabric(); len(fab) > 0 {
before := n.NVLinkDomain()
for _, f := range fab {
for i := range n.GPUs {
if n.GPUs[i].Index == int(f.GetIndex()) {
n.GPUs[i].Fabric = fabricFrom(f.GetState(), f.GetStatus(), f.GetClusterUuid(), f.GetCliqueId(), f.GetBandwidth())
}
}
}
if after := n.NVLinkDomain(); after != before {
s.log.Warn("a node's NVLink domain changed", "node", nodeID, "from", before, "to", after)
}
}
if m := hb.GetImex(); m != nil {
n.IMEX = discovery.IMEX{State: m.GetState(), Channels: int(m.GetChannels())}
}
// A port went down, or came back. Placement reads the fabrics, so this is
// what stops a node being offered multi-node work it can no longer carry.
if hb.GetFabricsChanged() {
@@ -74,6 +74,9 @@ type Collector struct {
ibRate *prometheus.Desc
ibCounter *prometheus.Desc
gpuNVLJoined *prometheus.Desc
imexUp *prometheus.Desc
fabricSMs *prometheus.Desc
fabricMaster *prometheus.Desc
fabricProblems *prometheus.Desc
@@ -166,6 +169,11 @@ func New(src Source) *Collector {
ibRate: prometheus.NewDesc("llmgw_ib_port_rate_gbps",
"A port's negotiated rate. A link that trained below its usual speed is a "+
"cable or transceiver to look at.", []string{"node", "hca", "port"}, nil),
gpuNVLJoined: prometheus.NewDesc("llmgw_gpu_nvlink_joined",
"1 when a GPU is joined to its multi-node NVLink domain, 0 when it reports a fabric it has not "+
"joined. Absent on cards with no multi-node fabric.", []string{"node", "gpu", "domain", "state"}, nil),
imexUp: prometheus.NewDesc("llmgw_imex_up",
"1 when nvidia-imex is active with a channel, on nodes whose GPUs report a fabric.", []string{"node"}, nil),
fabricSMs: prometheus.NewDesc("llmgw_fabric_subnet_managers",
"Subnet managers a physical fabric lists, by state. No standby means the fabric "+
"cannot recover from a change if its master stops.", []string{"fabric", "physical", "state"}, nil),
@@ -189,6 +197,7 @@ func (c *Collector) Describe(ch chan<- *prometheus.Desc) {
c.nodeDiskFree, c.nodeDiskSize,
c.nodeCPU, c.nodeCPUs, c.nodeLoad, c.nodeMemUsed, c.nodeMemAvail, c.shardCPU, c.shardMemory,
c.ibActive, c.ibRate, c.ibCounter, c.fabricSMs, c.fabricMaster, c.fabricProblems,
c.gpuNVLJoined, c.imexUp,
c.gpuMemoryTotal, c.gpuMemoryUsed, c.gpuUtil, c.gpuTemp, c.gpuHealthy,
c.gpuNVLink, c.gpuECC,
c.replicaState, c.replicaReady, c.replicaIdle,
@@ -263,6 +272,28 @@ func (c *Collector) Collect(ch chan<- prometheus.Metric) {
}
}
fabricGPUs := 0
for _, g := range n.GPUs {
fb := g.Fabric
if fb == nil || strings.EqualFold(fb.State, "Not Supported") {
continue
}
fabricGPUs++
joined := 0.0
if fb.Domain() != "" {
joined = 1
}
ch <- prometheus.MustNewConstMetric(c.gpuNVLJoined, prometheus.GaugeValue, joined,
n.ID, strconv.Itoa(g.Index), fb.Domain(), fb.State)
}
if fabricGPUs > 0 {
up := 0.0
if n.IMEX.State == "active" && n.IMEX.Channels > 0 {
up = 1
}
ch <- prometheus.MustNewConstMetric(c.imexUp, prometheus.GaugeValue, up, n.ID)
}
byClass := map[string]int{}
for _, g := range n.GPUs {
class := g.Class()
+18
View File
@@ -120,6 +120,24 @@ type Node struct {
IBPorts []IBPort `json:"-"`
// IBAt is when IBPorts was read, for the rates.
IBAt time.Time `json:"-"`
// IMEX is the service a multi-node NVLink domain needs on this node.
IMEX discovery.IMEX `json:"imex"`
}
// NVLinkDomain is the multi-node NVLink domain this node's GPUs are joined
// to, when every one of them is joined to the same. A node half joined — a
// card still coming up, or one that fell out — is in none: a replica that
// counts on its NVLink would find a card that has none.
func (n Node) NVLinkDomain() string {
d := ""
for _, g := range n.GPUs {
gd := g.Fabric.Domain()
if gd == "" || (d != "" && gd != d) {
return ""
}
d = gd
}
return d
}
// IBPort is one port's state, identity and cumulative counters.
+172
View File
@@ -0,0 +1,172 @@
// Package nvlink works out the cluster's multi-node NVLink domains — an NVL72
// rack, an NVLink Switch system — from what each GPU reports, and what keeps
// one from being usable.
//
// A domain is a scale-up island: its GPUs reach each other over NVLink
// whatever node they are in, so a tensor-parallel group can span its nodes
// at NVLink speed. What it needs to be one is every card joined by the
// fabric manager and IMEX running on every node; either missing, and a
// replica placed across it fails or crawls.
package nvlink
import (
"fmt"
"sort"
"strings"
"gitea.mynodes.xyz/ilker/llmgw/internal/control/inventory"
)
// Level is how much a problem matters.
type Level string
// The levels.
const (
Fail Level = "fail"
Warn Level = "warn"
)
// Problem is one thing wrong with a domain.
type Problem struct {
Level Level `json:"level"`
Domain string `json:"domain"`
Message string `json:"message"`
}
// GPU is one card as its domain sees it.
type GPU struct {
Index int `json:"index"`
Name string `json:"name"`
State string `json:"state"`
Status string `json:"status,omitempty"`
Bandwidth string `json:"bandwidth,omitempty"`
Joined bool `json:"joined"`
}
// Member is one node of a domain.
type Member struct {
Node string `json:"node"`
GPUs []GPU `json:"gpus"`
// Usable is every card joined and IMEX up: what placement requires.
Usable bool `json:"usable"`
IMEX string `json:"imex"`
IMEXChannels int `json:"imex_channels"`
}
// Domain is one multi-node NVLink domain.
type Domain struct {
Key string `json:"key"`
ClusterUUID string `json:"cluster_uuid"`
CliqueID string `json:"clique_id"`
Members []Member `json:"members"`
GPUs int `json:"gpus"`
Joined int `json:"joined"`
Usable int `json:"usable_nodes"`
}
// Report is every domain and what is wrong with them.
type Report struct {
Domains []Domain `json:"domains"`
Problems []Problem `json:"problems"`
}
// Analyze reads every node's GPUs.
func Analyze(nodes []inventory.Node) Report {
var r Report
by := map[string]*Domain{}
for _, n := range nodes {
// The domain a node is in is the one any of its cards joined; a card
// still joining has no domain of its own yet, and is counted against
// the node's.
key := ""
for _, g := range n.GPUs {
if d := g.Fabric.Domain(); d != "" {
key = d
break
}
}
if key == "" {
continue
}
d := by[key]
if d == nil {
uuid, clique, _ := strings.Cut(key, "/")
d = &Domain{Key: key, ClusterUUID: uuid, CliqueID: clique}
by[key] = d
}
m := Member{Node: n.ID, IMEX: n.IMEX.State, IMEXChannels: n.IMEX.Channels}
joined := 0
for _, g := range n.GPUs {
v := GPU{Index: g.Index, Name: g.Name}
if f := g.Fabric; f != nil {
v.State, v.Status, v.Bandwidth = f.State, f.Status, f.Bandwidth
v.Joined = f.Domain() == key
}
if v.Joined {
joined++
} else {
why := v.State
if why == "" {
why = "no fabric"
}
if v.Status != "" && !strings.EqualFold(v.Status, "Success") {
why += ": " + v.Status
}
r.Problems = append(r.Problems, Problem{Level: Fail, Domain: key, Message: fmt.Sprintf(
"%s GPU %d is not joined to its domain (%s)", n.ID, g.Index, why)})
}
if v.Bandwidth != "" && !strings.EqualFold(v.Bandwidth, "Full") {
r.Problems = append(r.Problems, Problem{Level: Warn, Domain: key, Message: fmt.Sprintf(
"%s GPU %d has %s NVLink bandwidth", n.ID, g.Index, strings.ToLower(v.Bandwidth))})
}
m.GPUs = append(m.GPUs, v)
}
imexOK := n.IMEX.State == "active" && n.IMEX.Channels > 0
switch {
case n.IMEX.State != "active":
state := n.IMEX.State
if state == "" {
state = "not installed"
}
r.Problems = append(r.Problems, Problem{Level: Fail, Domain: key, Message: fmt.Sprintf(
"%s: nvidia-imex is %s — without it the domain's GPUs cannot share memory across nodes",
n.ID, state)})
case n.IMEX.Channels == 0:
r.Problems = append(r.Problems, Problem{Level: Fail, Domain: key, Message: fmt.Sprintf(
"%s: nvidia-imex runs but there is no IMEX channel for an engine to open (/dev/nvidia-caps-imex-channels)", n.ID)})
}
m.Usable = joined == len(n.GPUs) && joined > 0 && imexOK
if joined > 0 && joined < len(n.GPUs) {
r.Problems = append(r.Problems, Problem{Level: Fail, Domain: key, Message: fmt.Sprintf(
"%s is half joined (%d of %d GPUs), so no replica spans the domain through it", n.ID, joined, len(n.GPUs))})
}
d.GPUs += len(n.GPUs)
d.Joined += joined
if m.Usable {
d.Usable++
}
d.Members = append(d.Members, m)
}
for _, d := range by {
sort.Slice(d.Members, func(i, j int) bool { return d.Members[i].Node < d.Members[j].Node })
r.Domains = append(r.Domains, *d)
}
sort.Slice(r.Domains, func(i, j int) bool { return r.Domains[i].Key < r.Domains[j].Key })
sort.SliceStable(r.Problems, func(i, j int) bool {
if r.Problems[i].Level != r.Problems[j].Level {
return r.Problems[i].Level == Fail
}
return r.Problems[i].Message < r.Problems[j].Message
})
return r
}
// Short is a domain's key as a person reads it: the cluster's first group
// and the clique.
func Short(key string) string {
uuid, clique, _ := strings.Cut(key, "/")
if i := strings.IndexByte(uuid, '-'); i > 0 {
uuid = uuid[:i]
}
return uuid + "·" + clique
}
+55
View File
@@ -0,0 +1,55 @@
package nvlink
import (
"strings"
"testing"
"gitea.mynodes.xyz/ilker/llmgw/internal/agent/discovery"
"gitea.mynodes.xyz/ilker/llmgw/internal/control/inventory"
)
func tray(id string, gpus int, state string) inventory.Node {
n := inventory.Node{ID: id, IMEX: discovery.IMEX{State: "active", Channels: 1}}
for i := range gpus {
n.GPUs = append(n.GPUs, discovery.GPU{Index: i, Name: "NVIDIA GB200", Fabric: &discovery.GPUFabric{
State: state, Status: "Success", ClusterUUID: "76b1ad76-7c24-4b6b-9f6c-2a1e0d8c11aa", CliqueID: "1", Bandwidth: "Full"}})
}
return n
}
func has(t *testing.T, r Report, l Level, s string) {
t.Helper()
for _, p := range r.Problems {
if p.Level == l && strings.Contains(p.Message, s) {
return
}
}
t.Errorf("no %s problem with %q in %+v", l, s, r.Problems)
}
func TestARackOfJoinedTraysIsOneUsableDomain(t *testing.T) {
r := Analyze([]inventory.Node{tray("t1", 4, "Completed"), tray("t2", 4, "Completed"), {ID: "cpu"}})
if len(r.Domains) != 1 || r.Domains[0].Usable != 2 || r.Domains[0].Joined != 8 || len(r.Problems) != 0 {
t.Errorf("report = %+v", r)
}
if Short(r.Domains[0].Key) != "76b1ad76·1" {
t.Errorf("short = %q", Short(r.Domains[0].Key))
}
}
func TestWhatKeepsATrayOutIsSaid(t *testing.T) {
half := tray("t2", 4, "Completed")
half.GPUs[3].Fabric = &discovery.GPUFabric{State: "In Progress"}
noIMEX := tray("t3", 4, "Completed")
noIMEX.IMEX = discovery.IMEX{State: "inactive"}
slow := tray("t4", 4, "Completed")
slow.GPUs[0].Fabric.Bandwidth = "Degraded"
r := Analyze([]inventory.Node{tray("t1", 4, "Completed"), half, noIMEX, slow})
has(t, r, Fail, "t2 GPU 3 is not joined to its domain (In Progress)")
has(t, r, Fail, "t2 is half joined")
has(t, r, Fail, "t3: nvidia-imex is inactive")
has(t, r, Warn, "t4 GPU 0 has degraded NVLink bandwidth")
if r.Domains[0].Usable != 2 {
t.Errorf("usable = %d, want t1 and t4", r.Domains[0].Usable)
}
}
+68 -1
View File
@@ -239,13 +239,20 @@ func (s *Scheduler) placeMultiNode(p Profile, avoid []string) (*Placement, []Rej
var rejections []Rejection
fabrics := s.inv.Fabrics()
if len(fabrics) == 0 {
if len(fabrics) == 0 && !s.anyNVLinkDomain() {
return nil, []Rejection{{
Profile: p.Name,
Reason: fmt.Sprintf("needs %d nodes on one fabric, but no node reports a fabric", p.Nodes),
}}
}
// A multi-node NVLink domain first: a replica inside one runs tensor
// parallelism across its nodes at NVLink speed, which no InfiniBand
// fabric matches.
if pl := s.placeOnNVLink(p, avoid, &rejections); pl != nil {
return pl, rejections
}
canon := inventory.CanonicalSMs(s.inv.Nodes())
for _, fabric := range fabrics {
all := s.eligibleNodes(p, fabric, avoid, &rejections)
@@ -262,6 +269,66 @@ func (s *Scheduler) placeMultiNode(p Profile, avoid []string) (*Placement, []Rej
return nil, rejections
}
// NVLinkFabric is how a placement names a multi-node NVLink domain in its
// Fabric, so the shards on it are told to use it.
const NVLinkFabric = "nvl:"
func (s *Scheduler) anyNVLinkDomain() bool {
for _, n := range s.inv.Nodes() {
if n.NVLinkDomain() != "" {
return true
}
}
return false
}
// placeOnNVLink forms a replica inside one multi-node NVLink domain, from
// nodes whose every card is joined to it and whose IMEX is up: a node
// missing either would leave the replica a card, or its memory, short.
func (s *Scheduler) placeOnNVLink(p Profile, avoid []string, rejections *[]Rejection) *Placement {
var none []Rejection
cands := s.eligibleNodes(p, "", avoid, &none)
by := map[string][]candidate{}
for _, c := range cands {
d := c.node.NVLinkDomain()
if d == "" {
continue
}
if c.node.IMEX.State != "active" || c.node.IMEX.Channels == 0 {
*rejections = append(*rejections, Rejection{Profile: p.Name, Pool: NVLinkFabric + shortDomain(d), Node: c.node.ID,
Reason: "in an NVLink domain, but nvidia-imex is not up"})
continue
}
by[d] = append(by[d], c)
}
keys := make([]string, 0, len(by))
for k := range by {
keys = append(keys, k)
}
sort.Slice(keys, func(i, j int) bool {
if len(by[keys[i]]) != len(by[keys[j]]) {
return len(by[keys[i]]) > len(by[keys[j]])
}
return keys[i] < keys[j]
})
for _, d := range keys {
pool := NVLinkFabric + shortDomain(d) + "/" + p.classLabel()
if pl := s.placeOnFabric(p, NVLinkFabric+d, pool, by[d], rejections); pl != nil {
return pl
}
}
return nil
}
// shortDomain is an NVLink domain as a pool name carries it.
func shortDomain(d string) string {
uuid, clique, _ := strings.Cut(d, "/")
if i := strings.IndexByte(uuid, '-'); i > 0 {
uuid = uuid[:i]
}
return uuid + "." + clique
}
// fabricGroup is the candidates on one physical fabric.
type fabricGroup struct {
key string
@@ -1018,3 +1018,56 @@ func TestAReplicaStaysUnderOneLeafAndOffDegradedCables(t *testing.T) {
t.Errorf("placed on %v, want the two good nodes under leaf-b", got)
}
}
// inDomain joins every card of a node to one multi-node NVLink domain, with
// IMEX up.
func inDomain(n inventory.Node, clique string) inventory.Node {
for i := range n.GPUs {
n.GPUs[i].Fabric = &discovery.GPUFabric{State: "Completed", ClusterUUID: "76b1ad76-7c24-4b6b-9f6c-2a1e0d8c11aa", CliqueID: clique}
}
n.IMEX = discovery.IMEX{State: "active", Channels: 1}
return n
}
func TestAReplicaSpansNodesOverNVLinkBeforeInfiniBand(t *testing.T) {
// Four trays on one InfiniBand fabric, two of them in an NVLink domain:
// the pair is placed across the domain, where tensor parallelism between
// nodes runs at NVLink speed, and given no adapters to use.
st := store(
onSM(node("ib1", "h100-80g", 4, "ib-a"), "0x01"), onSM(node("ib2", "h100-80g", 4, "ib-a"), "0x01"),
inDomain(onSM(node("nv1", "h100-80g", 4, "ib-a"), "0x01"), "7"),
inDomain(onSM(node("nv2", "h100-80g", 4, "ib-a"), "0x01"), "7"),
)
res := New(st).Place([]Profile{{Name: "tp8", GPUClass: "h100-80g", GPUsPerNode: 4, Nodes: 2}}, nil)
if res.Placement == nil {
t.Fatalf("no placement:\n%s", res.Explain())
}
if !strings.HasPrefix(res.Placement.Fabric, NVLinkFabric) || !strings.HasPrefix(res.Placement.Pool, "nvl:76b1ad76.7/") {
t.Errorf("fabric %q pool %q", res.Placement.Fabric, res.Placement.Pool)
}
for _, a := range res.Placement.Nodes {
if !strings.HasPrefix(a.NodeID, "nv") || len(a.HCAs) != 0 {
t.Errorf("assignment %+v", a)
}
}
}
func TestANodeWithoutIMEXIsNotCountedInItsDomain(t *testing.T) {
a := inDomain(onSM(node("nv1", "h100-80g", 4, "ib-a"), "0x01"), "7")
b := inDomain(onSM(node("nv2", "h100-80g", 4, "ib-a"), "0x01"), "7")
b.IMEX.State = "inactive"
res := New(store(a, b)).Place([]Profile{{Name: "tp8", GPUClass: "h100-80g", GPUsPerNode: 4, Nodes: 2}}, nil)
// Not across the domain; InfiniBand still carries it.
if res.Placement == nil || strings.HasPrefix(res.Placement.Fabric, NVLinkFabric) {
t.Fatalf("placement %+v\n%s", res.Placement, res.Explain())
}
// With no InfiniBand to fall back on, the refusal says why.
a2 := inDomain(node("nv1", "h100-80g", 4), "7")
b2 := inDomain(node("nv2", "h100-80g", 4), "7")
b2.IMEX.State = "inactive"
res = New(store(a2, b2)).Place([]Profile{{Name: "tp8", GPUClass: "h100-80g", GPUsPerNode: 4, Nodes: 2}}, nil)
if res.Placement != nil || !strings.Contains(res.Explain(), "nvidia-imex is not up") {
t.Errorf("the explanation should say why the domain was passed over:\n%s", res.Explain())
}
}