Merge branch '13579-keepstore-config-doc' refs #13579 refs #13580

[arvados.git] / lib / dispatchcloud / node_size.go

diff --git a/lib/dispatchcloud/node_size.go b/lib/dispatchcloud/node_size.go
index 41c6ff425121515101bdaf1d7aa5b4004acef57d..b5fd0262a8e430ce9cdaac4c58c5078fe7e1e314 100644 (file)
--- a/lib/dispatchcloud/node_size.go
+++ b/lib/dispatchcloud/node_size.go
@@ -5,10 +5,10 @@
 package dispatchcloud
 
 import (
-       "bytes"
        "errors"
        "log"
        "os/exec"
+       "sort"
        "strings"
        "time"
 
@@ -16,11 +16,17 @@ import (
 )
 
 var (
-       ErrConstraintsNotSatisfiable  = errors.New("constraints not satisfiable by any configured instance type")
        ErrInstanceTypesNotConfigured = errors.New("site configuration does not list any instance types")
        discountConfiguredRAMPercent  = 5
 )
 
+// ConstraintsNotSatisfiableError includes a list of available instance types
+// to be reported back to the user.
+type ConstraintsNotSatisfiableError struct {
+       error
+       AvailableTypes []arvados.InstanceType
+}
+
 // ChooseInstanceType returns the cheapest available
 // arvados.InstanceType big enough to run ctr.
 func ChooseInstanceType(cc *arvados.Cluster, ctr *arvados.Container) (best arvados.InstanceType, err error) {
@@ -41,13 +47,22 @@ func ChooseInstanceType(cc *arvados.Cluster, ctr *arvados.Container) (best arvad
        needRAM := ctr.RuntimeConstraints.RAM + ctr.RuntimeConstraints.KeepCacheRAM
        needRAM = (needRAM * 100) / int64(100-discountConfiguredRAMPercent)
 
-       err = ErrConstraintsNotSatisfiable
+       availableTypes := make([]arvados.InstanceType, len(cc.InstanceTypes))
+       copy(availableTypes, cc.InstanceTypes)
+       sort.Slice(availableTypes, func(a, b int) bool {
+               return availableTypes[a].Price < availableTypes[b].Price
+       })
+       err = ConstraintsNotSatisfiableError{
+               errors.New("constraints not satisfiable by any configured instance type"),
+               availableTypes,
+       }
        for _, it := range cc.InstanceTypes {
                switch {
                case err == nil && it.Price > best.Price:
                case it.Scratch < needScratch:
                case it.RAM < needRAM:
                case it.VCPUs < needVCPUs:
+               case it.Preemptable != ctr.SchedulingParameters.Preemptable:
                case it.Price == best.Price && (it.RAM < best.RAM || it.VCPUs < best.VCPUs):
                        // Equal price, but worse specs
                default:
@@ -88,35 +103,31 @@ func SlurmNodeTypeFeatureKludge(cc *arvados.Cluster) {
        }
        for {
                slurmKludge(features)
-               time.Sleep(time.Minute)
+               time.Sleep(2 * time.Second)
        }
 }
 
-var (
-       slurmDummyNode     = "compute0"
-       slurmErrBadFeature = "Invalid feature"
-       slurmErrNoNodes    = "node configuration is not available"
-)
+const slurmDummyNode = "compute0"
 
 func slurmKludge(features []string) {
-       cmd := exec.Command("srun", "--test-only", "--constraint="+strings.Join(features, "&"), "false")
-       out, err := cmd.CombinedOutput()
-       switch {
-       case err == nil || bytes.Contains(out, []byte(slurmErrNoNodes)):
-               // Evidently our node-type feature names are all valid.
+       allFeatures := strings.Join(features, ",")
 
-       case bytes.Contains(out, []byte(slurmErrBadFeature)):
-               log.Printf("temporarily configuring node %q with all node type features", slurmDummyNode)
-               for _, nodeFeatures := range []string{strings.Join(features, ","), ""} {
-                       cmd = exec.Command("scontrol", "update", "NodeName="+slurmDummyNode, "Features="+nodeFeatures)
-                       log.Printf("running: %q %q", cmd.Path, cmd.Args)
-                       out, err := cmd.CombinedOutput()
-                       if err != nil {
-                               log.Printf("error: scontrol: %s (output was %q)", err, out)
-                       }
-               }
+       cmd := exec.Command("sinfo", "--nodes="+slurmDummyNode, "--format=%f", "--noheader")
+       out, err := cmd.CombinedOutput()
+       if err != nil {
+               log.Printf("running %q %q: %s (output was %q)", cmd.Path, cmd.Args, err, out)
+               return
+       }
+       if string(out) == allFeatures+"\n" {
+               // Already configured correctly, nothing to do.
+               return
+       }
 
-       default:
-               log.Printf("warning: expected srun error %q, %q, or success, but output was %q", slurmErrBadFeature, slurmErrNoNodes, out)
+       log.Printf("configuring node %q with all node type features", slurmDummyNode)
+       cmd = exec.Command("scontrol", "update", "NodeName="+slurmDummyNode, "Features="+allFeatures)
+       log.Printf("running: %q %q", cmd.Path, cmd.Args)
+       out, err = cmd.CombinedOutput()
+       if err != nil {
+               log.Printf("error: scontrol: %s (output was %q)", err, out)
        }
 }