work on socket scoped metrics

This commit is contained in:
Lou Knauer 2022-01-12 13:03:01 +01:00
parent f185d12078
commit 98c4de65a7
5 changed files with 198 additions and 115 deletions

View File

@ -15,3 +15,31 @@ type MetricDataRepository struct {
Url string `json:"url"`
Token string `json:"token"`
}
// Return a list of socket IDs given a list of hwthread IDs.
// Even if just one hwthread is in that socket, add it to the list.
// If no hwthreads other than those in the argument list are assigned to
// one of the sockets in the first return value, return true as the second value.
// TODO: Optimize this, there must be a more efficient way/algorithm.
func (topo *Topology) GetSockets(hwthreads []int) (sockets []int, exclusive bool) {
socketsMap := map[int]int{}
for _, hwthread := range hwthreads {
for socket, hwthreadsInSocket := range topo.Socket {
for _, hwthreadInSocket := range hwthreadsInSocket {
if hwthread == hwthreadInSocket {
socketsMap[socket] += 1
}
}
}
}
exclusive = true
hwthreadsPerSocket := len(topo.Node) / len(topo.Socket)
sockets = make([]int, 0, len(socketsMap))
for socket, count := range socketsMap {
sockets = append(sockets, socket)
exclusive = exclusive && count == hwthreadsPerSocket
}
return sockets, exclusive
}

View File

@ -144,7 +144,7 @@ func ArchiveJob(job *schema.Job, ctx context.Context) (*schema.JobMeta, error) {
return nil, err
}
if err := calcStatisticsSeries(job, jobData); err != nil {
if err := calcStatisticsSeries(job, jobData, 7); err != nil {
return nil, err
}
@ -221,14 +221,14 @@ func ArchiveJob(job *schema.Job, ctx context.Context) (*schema.JobMeta, error) {
}
// Add statisticsSeries fields
func calcStatisticsSeries(job *schema.Job, jobData schema.JobData) error {
func calcStatisticsSeries(job *schema.Job, jobData schema.JobData, maxSeries int) error {
for _, scopes := range jobData {
for _, jobMetric := range scopes {
if jobMetric.StatisticsSeries != nil {
continue
}
if len(jobMetric.Series) < 5 {
if len(jobMetric.Series) <= maxSeries {
continue
}

View File

@ -7,7 +7,6 @@ import (
"encoding/json"
"errors"
"fmt"
"log"
"net/http"
"strconv"
"time"
@ -27,6 +26,15 @@ type ApiRequestBody struct {
Selectors [][]string `json:"selectors"`
}
type ApiQuery struct {
Metric string `json:"metric"`
Hostname string `json:"hostname"`
Type *string `json:"type,omitempty"`
TypeIds []string `json:"type-ids,omitempty"`
SubType *string `json:"subtype,omitempty"`
SubTypeIds []string `json:"subtype-ids,omitempty"`
}
type ApiMetricData struct {
Error *string `json:"error"`
From int64 `json:"from"`
@ -50,6 +58,9 @@ type ApiStatsData struct {
func (ccms *CCMetricStore) Init(url, token string) error {
ccms.url = url
ccms.jwt = token
ccms.client = http.Client{
Timeout: 5 * time.Second,
}
return nil
}
@ -82,17 +93,6 @@ func (ccms *CCMetricStore) doRequest(job *schema.Job, suffix string, metrics []s
}
func (ccms *CCMetricStore) LoadData(job *schema.Job, metrics []string, scopes []schema.MetricScope, ctx context.Context) (schema.JobData, error) {
// log.Printf("job: %#v", job)
type ApiQuery struct {
Metric string `json:"metric"`
Hostname string `json:"hostname"`
Type *string `json:"type,omitempty"`
TypeIds []string `json:"type-ids,omitempty"`
SubType *string `json:"subtype,omitempty"`
SubTypeIds []string `json:"subtype-ids,omitempty"`
}
type ApiQueryRequest struct {
Cluster string `json:"cluster"`
From int64 `json:"from"`
@ -105,101 +105,18 @@ func (ccms *CCMetricStore) LoadData(job *schema.Job, metrics []string, scopes []
Query *ApiQuery `json:"query"`
}
queries, scopeForMetric, err := ccms.buildQueries(job, metrics, scopes)
if err != nil {
return nil, err
}
reqBody := ApiQueryRequest{
Cluster: job.Cluster,
From: job.StartTime.Unix(),
To: job.StartTime.Add(time.Duration(job.Duration) * time.Second).Unix(),
Queries: make([]ApiQuery, 0),
Queries: queries,
}
if len(scopes) != 1 {
return nil, errors.New("todo: support more than one scope in a query")
}
topology := config.GetPartition(job.Cluster, job.Partition).Topology
scopeForMetric := map[string]schema.MetricScope{}
for _, metric := range metrics {
mc := config.GetMetricConfig(job.Cluster, metric)
if mc == nil {
// return nil, fmt.Errorf("metric '%s' is not specified for cluster '%s'", metric, job.Cluster)
log.Printf("metric '%s' is not specified for cluster '%s'", metric, job.Cluster)
continue
}
nativeScope, requestedScope := mc.Scope, scopes[0]
// case 1: A metric is requested at node scope with a native scope of node as well
// case 2: A metric is requested at node scope and node is exclusive
// case 3: A metric has native scope node
if (nativeScope == requestedScope && nativeScope == schema.MetricScopeNode) ||
(job.Exclusive == 1 && requestedScope == schema.MetricScopeNode) ||
(nativeScope == schema.MetricScopeNode) {
nodes := map[string]bool{}
for _, resource := range job.Resources {
nodes[resource.Hostname] = true
}
for node := range nodes {
reqBody.Queries = append(reqBody.Queries, ApiQuery{
Metric: metric,
Hostname: node,
})
}
scopeForMetric[metric] = schema.MetricScopeNode
continue
}
// case: Read a metric at hwthread scope with native scope hwthread
if nativeScope == requestedScope && nativeScope == schema.MetricScopeHWThread && job.NumNodes == 1 {
hwthreads := job.Resources[0].HWThreads
if hwthreads == nil {
hwthreads = topology.Node
}
t := "cpu" // TODO/FIXME: inconsistency between cc-metric-collector and ClusterCockpit
for _, hwthread := range hwthreads {
reqBody.Queries = append(reqBody.Queries, ApiQuery{
Metric: metric,
Hostname: job.Resources[0].Hostname,
Type: &t,
TypeIds: []string{strconv.Itoa(hwthread)},
})
}
scopeForMetric[metric] = schema.MetricScopeHWThread
continue
}
// case: A metric is requested at node scope, has a hwthread scope and node is not exclusive and runs on a single node
if requestedScope == schema.MetricScopeNode && nativeScope == schema.MetricScopeHWThread && job.Exclusive != 1 && job.NumNodes == 1 {
hwthreads := job.Resources[0].HWThreads
if hwthreads == nil {
hwthreads = topology.Node
}
t := "cpu" // TODO/FIXME: inconsistency between cc-metric-collector and ClusterCockpit
ids := make([]string, 0, len(hwthreads))
for _, hwthread := range hwthreads {
ids = append(ids, strconv.Itoa(hwthread))
}
reqBody.Queries = append(reqBody.Queries, ApiQuery{
Metric: metric,
Hostname: job.Resources[0].Hostname,
Type: &t,
TypeIds: ids,
})
scopeForMetric[metric] = schema.MetricScopeNode
continue
}
// TODO: Job teilt sich knoten und metric native scope ist kleiner als node
panic("todo")
}
// log.Printf("query: %#v", reqBody)
buf := &bytes.Buffer{}
if err := json.NewEncoder(buf).Encode(reqBody); err != nil {
return nil, err
@ -281,6 +198,142 @@ func (ccms *CCMetricStore) LoadData(job *schema.Job, metrics []string, scopes []
return jobData, nil
}
var (
cpuString = string(schema.MetricScopeCpu)
socketString = string(schema.MetricScopeSocket)
acceleratorString = string(schema.MetricScopeAccelerator)
)
func (ccms *CCMetricStore) buildQueries(job *schema.Job, metrics []string, scopes []schema.MetricScope) ([]ApiQuery, map[string]schema.MetricScope, error) {
queries := make([]ApiQuery, 0, len(metrics)*len(scopes)*len(job.Resources))
assignedScopes := make(map[string]schema.MetricScope, len(metrics))
topology := config.GetPartition(job.Cluster, job.Partition).Topology
if len(scopes) != 1 {
return nil, nil, errors.New("todo: support more than one scope in a query")
}
_ = topology
for _, metric := range metrics {
mc := config.GetMetricConfig(job.Cluster, metric)
if mc == nil {
// return nil, fmt.Errorf("metric '%s' is not specified for cluster '%s'", metric, job.Cluster)
// log.Printf("metric '%s' is not specified for cluster '%s'", metric, job.Cluster)
continue
}
nativeScope, requestedScope := mc.Scope, scopes[0]
// case 1: A metric is requested at node scope with a native scope of node as well
// case 2: A metric is requested at node scope and node is exclusive
// case 3: A metric has native scope node
if (nativeScope == requestedScope && nativeScope == schema.MetricScopeNode) ||
(job.Exclusive == 1 && requestedScope == schema.MetricScopeNode) ||
(nativeScope == schema.MetricScopeNode) {
nodes := map[string]bool{}
for _, resource := range job.Resources {
nodes[resource.Hostname] = true
}
for node := range nodes {
queries = append(queries, ApiQuery{
Metric: metric,
Hostname: node,
})
}
assignedScopes[metric] = schema.MetricScopeNode
continue
}
// case: Read a metric at hwthread scope with native scope hwthread
if nativeScope == requestedScope && nativeScope == schema.MetricScopeHWThread && job.NumNodes == 1 {
hwthreads := job.Resources[0].HWThreads
if hwthreads == nil {
hwthreads = topology.Node
}
for _, hwthread := range hwthreads {
queries = append(queries, ApiQuery{
Metric: metric,
Hostname: job.Resources[0].Hostname,
Type: &cpuString, // TODO/FIXME: inconsistency between cc-metric-collector and ClusterCockpit
TypeIds: []string{strconv.Itoa(hwthread)},
})
}
assignedScopes[metric] = schema.MetricScopeHWThread
continue
}
// case: A metric is requested at node scope, has a hwthread scope and node is not exclusive and runs on a single node
if requestedScope == schema.MetricScopeNode && nativeScope == schema.MetricScopeHWThread && job.Exclusive != 1 && job.NumNodes == 1 {
hwthreads := job.Resources[0].HWThreads
if hwthreads == nil {
hwthreads = topology.Node
}
ids := make([]string, 0, len(hwthreads))
for _, hwthread := range hwthreads {
ids = append(ids, strconv.Itoa(hwthread))
}
queries = append(queries, ApiQuery{
Metric: metric,
Hostname: job.Resources[0].Hostname,
Type: &cpuString, // TODO/FIXME: inconsistency between cc-metric-collector and ClusterCockpit
TypeIds: ids,
})
assignedScopes[metric] = schema.MetricScopeNode
continue
}
// case: A metric of native scope socket is requested at any scope lower than node and runs on a single node
if requestedScope.LowerThan(schema.MetricScopeNode) && nativeScope == schema.MetricScopeSocket && job.NumNodes == 1 {
hwthreads := job.Resources[0].HWThreads
if hwthreads == nil {
hwthreads = topology.Node
}
sockets, _ := topology.GetSockets(hwthreads)
ids := make([]string, 0, len(sockets))
for _, socket := range sockets {
ids = append(ids, strconv.Itoa(socket))
}
queries = append(queries, ApiQuery{
Metric: metric,
Hostname: job.Resources[0].Hostname,
Type: &socketString,
TypeIds: ids,
})
assignedScopes[metric] = schema.MetricScopeNode
continue
}
// case: A metric of native scope accelerator is requested at a sub-node scope
if requestedScope.LowerThan(schema.MetricScopeNode) && nativeScope == schema.MetricScopeAccelerator {
for _, resource := range job.Resources {
for _, acc := range resource.Accelerators {
queries = append(queries, ApiQuery{
Metric: metric,
Hostname: job.Resources[0].Hostname,
Type: &acceleratorString,
TypeIds: []string{strconv.Itoa(acc)},
})
}
}
assignedScopes[metric] = schema.MetricScopeAccelerator
}
// TODO: Job teilt sich knoten und metric native scope ist kleiner als node
panic("todo")
}
return queries, assignedScopes, nil
}
func (ccms *CCMetricStore) LoadStats(job *schema.Job, metrics []string, ctx context.Context) (map[string]map[string]schema.MetricStatistics, error) {
res, err := ccms.doRequest(job, "stats", metrics, ctx)
if err != nil {

View File

@ -16,7 +16,7 @@ type MetricDataRepository interface {
// Return the JobData for the given job, only with the requested metrics.
LoadData(job *schema.Job, metrics []string, scopes []schema.MetricScope, ctx context.Context) (schema.JobData, error)
// Return a map of metrics to a map of nodes to the metric statistics of the job.
// Return a map of metrics to a map of nodes to the metric statistics of the job. node scope assumed for now.
LoadStats(job *schema.Job, metrics []string, ctx context.Context) (map[string]map[string]schema.MetricStatistics, error)
// Return a map of nodes to a map of metrics to the data for the requested time.
@ -68,7 +68,7 @@ func LoadData(job *schema.Job, metrics []string, scopes []schema.MetricScope, ct
return nil, err
}
calcStatisticsSeries(job, data)
calcStatisticsSeries(job, data, 7)
return data, nil
}
@ -122,6 +122,7 @@ func LoadAverages(job *schema.Job, metrics []string, data [][]schema.Float, ctx
return nil
}
// Used for the node/system view. Returns a map of nodes to a map of metrics (at node scope).
func LoadNodeData(clusterId string, metrics, nodes []string, from, to int64, ctx context.Context) (map[string]map[string][]schema.Float, error) {
repo, ok := metricDataRepos[clusterId]
if !ok {

View File

@ -42,22 +42,23 @@ const (
MetricScopeSocket MetricScope = "socket"
MetricScopeCpu MetricScope = "cpu"
MetricScopeHWThread MetricScope = "hwthread"
MetricScopeAccelerator MetricScope = "accelerator"
)
var metricScopeGranularity map[MetricScope]int = map[MetricScope]int{
MetricScopeNode: 1,
MetricScopeSocket: 2,
MetricScopeCpu: 3,
MetricScopeHWThread: 4,
MetricScopeNode: 10,
MetricScopeSocket: 5,
MetricScopeCpu: 2,
MetricScopeHWThread: 1,
MetricScopeAccelerator: 5, // Special/Randomly choosen
}
func (e *MetricScope) MaxGranularity(other MetricScope) MetricScope {
func (e *MetricScope) LowerThan(other MetricScope) bool {
a := metricScopeGranularity[*e]
b := metricScopeGranularity[other]
if a < b {
return *e
}
return other
return a < b
}
func (e *MetricScope) UnmarshalGQL(v interface{}) error {