X-Git-Url: https://git.arvados.org/arvados.git/blobdiff_plain/596884ea30c551d7c2699b2cf3fdea2cede3101a..b205525d0b7c7b9042513fe77d2e8061534208ae:/lib/dispatchcloud/node_size.go diff --git a/lib/dispatchcloud/node_size.go b/lib/dispatchcloud/node_size.go index 55e9d1e793..e77c862b36 100644 --- a/lib/dispatchcloud/node_size.go +++ b/lib/dispatchcloud/node_size.go @@ -6,19 +6,14 @@ package dispatchcloud import ( "errors" - "log" - "os/exec" "sort" - "strings" - "time" "git.curoverse.com/arvados.git/sdk/go/arvados" ) -var ( - ErrInstanceTypesNotConfigured = errors.New("site configuration does not list any instance types") - discountConfiguredRAMPercent = 5 -) +var ErrInstanceTypesNotConfigured = errors.New("site configuration does not list any instance types") + +var discountConfiguredRAMPercent = 5 // ConstraintsNotSatisfiableError includes a list of available instance types // to be reported back to the user. @@ -51,8 +46,8 @@ func ChooseInstanceType(cc *arvados.Cluster, ctr *arvados.Container) (best arvad for _, it := range cc.InstanceTypes { switch { case ok && it.Price > best.Price: - case it.Scratch < needScratch: - case it.RAM < needRAM: + case int64(it.Scratch) < needScratch: + case int64(it.RAM) < needRAM: case it.VCPUs < needVCPUs: case it.Preemptible != ctr.SchedulingParameters.Preemptible: case it.Price == best.Price && (it.RAM < best.RAM || it.VCPUs < best.VCPUs): @@ -79,61 +74,3 @@ func ChooseInstanceType(cc *arvados.Cluster, ctr *arvados.Container) (best arvad } return } - -// SlurmNodeTypeFeatureKludge ensures SLURM accepts every instance -// type name as a valid feature name, even if no instances of that -// type have appeared yet. -// -// It takes advantage of some SLURM peculiarities: -// -// (1) A feature is valid after it has been offered by a node, even if -// it is no longer offered by any node. So, to make a feature name -// valid, we can add it to a dummy node ("compute0"), then remove it. -// -// (2) To test whether a set of feature names are valid without -// actually submitting a job, we can call srun --test-only with the -// desired features. -// -// SlurmNodeTypeFeatureKludge does a test-and-fix operation -// immediately, and then periodically, in case slurm restarts and -// forgets the list of valid features. It never returns (unless there -// are no node types configured, in which case it returns -// immediately), so it should generally be invoked with "go". -func SlurmNodeTypeFeatureKludge(cc *arvados.Cluster) { - if len(cc.InstanceTypes) == 0 { - return - } - var features []string - for _, it := range cc.InstanceTypes { - features = append(features, "instancetype="+it.Name) - } - for { - slurmKludge(features) - time.Sleep(2 * time.Second) - } -} - -const slurmDummyNode = "compute0" - -func slurmKludge(features []string) { - allFeatures := strings.Join(features, ",") - - 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 - } - - 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) - } -}