X-Git-Url: https://git.arvados.org/arvados.git/blobdiff_plain/4d27cabda122ebda4c8688580af9088c8732a074..2f66d4cc05e9442a9bb69969744d0750a02a1ed4:/lib/dispatchcloud/node_size.go diff --git a/lib/dispatchcloud/node_size.go b/lib/dispatchcloud/node_size.go index 339e042c1a..fd04860861 100644 --- a/lib/dispatchcloud/node_size.go +++ b/lib/dispatchcloud/node_size.go @@ -6,21 +6,16 @@ package dispatchcloud import ( "errors" - "log" - "os/exec" "regexp" "sort" "strconv" - "strings" - "time" - "git.curoverse.com/arvados.git/sdk/go/arvados" + "git.arvados.org/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. @@ -38,12 +33,10 @@ var pdhRegexp = regexp.MustCompile(`^[0-9a-f]{32}\+(\d+)$`) func estimateDockerImageSize(collectionPDH string) int64 { m := pdhRegexp.FindStringSubmatch(collectionPDH) if m == nil { - log.Printf("estimateDockerImageSize: '%v' did not match pdhRegexp, returning 0", collectionPDH) return 0 } n, err := strconv.ParseInt(m[1], 10, 64) if err != nil || n < 122 { - log.Printf("estimateDockerImageSize: short manifest %v or error (%v), returning 0", n, err) return 0 } // To avoid having to fetch the collection, take advantage of @@ -53,7 +46,7 @@ func estimateDockerImageSize(collectionPDH string) int64 { // the size of the manifest. // // Use the following heuristic: - // - Start with the length of the mainfest (n) + // - Start with the length of the manifest (n) // - Subtract 80 characters for the filename and file segment // - Divide by 42 to get the number of block identifiers ('hash\+size\ ' is 32+1+8+1) // - Assume each block is full, multiply by 64 MiB @@ -137,61 +130,3 @@ 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) - } -}