X-Git-Url: https://git.arvados.org/arvados.git/blobdiff_plain/cbf93e8d897448dbd52369afe89fef2392140ff1..HEAD:/lib/dispatchcloud/node_size.go diff --git a/lib/dispatchcloud/node_size.go b/lib/dispatchcloud/node_size.go index 339e042c1a..802bc65c28 100644 --- a/lib/dispatchcloud/node_size.go +++ b/lib/dispatchcloud/node_size.go @@ -6,21 +6,17 @@ package dispatchcloud import ( "errors" - "log" - "os/exec" + "math" "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 +34,10 @@ 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 +47,7 @@ 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 @@ -63,7 +57,7 @@ func estimateDockerImageSize(collectionPDH string) int64 { // EstimateScratchSpace estimates how much available disk space (in // bytes) is needed to run the container by summing the capacity // requested by 'tmp' mounts plus disk space required to load the -// Docker image. +// Docker image plus arv-mount block cache. func EstimateScratchSpace(ctr *arvados.Container) (needScratch int64) { for _, m := range ctr.Mounts { if m.Kind == "tmp" { @@ -87,15 +81,35 @@ func EstimateScratchSpace(ctr *arvados.Container) (needScratch int64) { // Now reserve space for the extracted image on disk. needScratch += dockerImageSize + // Now reserve space the arv-mount disk cache + needScratch += ctr.RuntimeConstraints.KeepCacheDisk + return } -// 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) { +// compareVersion returns true if vs1 < vs2, otherwise false +func versionLess(vs1 string, vs2 string) (bool, error) { + v1, err := strconv.ParseFloat(vs1, 64) + if err != nil { + return false, err + } + v2, err := strconv.ParseFloat(vs2, 64) + if err != nil { + return false, err + } + return v1 < v2, nil +} + +// ChooseInstanceType returns the arvados.InstanceTypes eligible to +// run ctr, i.e., those that have enough RAM, VCPUs, etc., and are not +// too expensive according to cluster configuration. +// +// The returned types are sorted with lower prices first. +// +// The error is non-nil if and only if the returned slice is empty. +func ChooseInstanceType(cc *arvados.Cluster, ctr *arvados.Container) ([]arvados.InstanceType, error) { if len(cc.InstanceTypes) == 0 { - err = ErrInstanceTypesNotConfigured - return + return nil, ErrInstanceTypesNotConfigured } needScratch := EstimateScratchSpace(ctr) @@ -103,25 +117,42 @@ func ChooseInstanceType(cc *arvados.Cluster, ctr *arvados.Container) (best arvad needVCPUs := ctr.RuntimeConstraints.VCPUs needRAM := ctr.RuntimeConstraints.RAM + ctr.RuntimeConstraints.KeepCacheRAM + needRAM += int64(cc.Containers.ReserveExtraRAM) + if cc.Containers.LocalKeepBlobBuffersPerVCPU > 0 { + // + 200 MiB for keepstore process + 10% for GOGC=10 + needRAM += 220 << 20 + // + 64 MiB for each blob buffer + 10% for GOGC=10 + needRAM += int64(cc.Containers.LocalKeepBlobBuffersPerVCPU * needVCPUs * (1 << 26) * 11 / 10) + } needRAM = (needRAM * 100) / int64(100-discountConfiguredRAMPercent) - ok := false + maxPriceFactor := math.Max(cc.Containers.MaximumPriceFactor, 1) + var types []arvados.InstanceType + var maxPrice float64 for _, it := range cc.InstanceTypes { + driverInsuff, driverErr := versionLess(it.CUDA.DriverVersion, ctr.RuntimeConstraints.CUDA.DriverVersion) + capabilityInsuff, capabilityErr := versionLess(it.CUDA.HardwareCapability, ctr.RuntimeConstraints.CUDA.HardwareCapability) + switch { - case ok && it.Price > best.Price: - 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): - // Equal price, but worse specs + // reasons to reject a node + case maxPrice > 0 && it.Price > maxPrice: // too expensive + case int64(it.Scratch) < needScratch: // insufficient scratch + case int64(it.RAM) < needRAM: // insufficient RAM + case it.VCPUs < needVCPUs: // insufficient VCPUs + case it.Preemptible != ctr.SchedulingParameters.Preemptible: // wrong preemptable setting + case it.CUDA.DeviceCount < ctr.RuntimeConstraints.CUDA.DeviceCount: // insufficient CUDA devices + case ctr.RuntimeConstraints.CUDA.DeviceCount > 0 && (driverInsuff || driverErr != nil): // insufficient driver version + case ctr.RuntimeConstraints.CUDA.DeviceCount > 0 && (capabilityInsuff || capabilityErr != nil): // insufficient hardware capability + // Don't select this node default: - // Lower price || (same price && better specs) - best = it - ok = true + // Didn't reject the node, so select it + types = append(types, it) + if newmax := it.Price * maxPriceFactor; newmax < maxPrice || maxPrice == 0 { + maxPrice = newmax + } } } - if !ok { + if len(types) == 0 { availableTypes := make([]arvados.InstanceType, 0, len(cc.InstanceTypes)) for _, t := range cc.InstanceTypes { availableTypes = append(availableTypes, t) @@ -129,69 +160,39 @@ func ChooseInstanceType(cc *arvados.Cluster, ctr *arvados.Container) (best arvad sort.Slice(availableTypes, func(a, b int) bool { return availableTypes[a].Price < availableTypes[b].Price }) - err = ConstraintsNotSatisfiableError{ + return nil, ConstraintsNotSatisfiableError{ errors.New("constraints not satisfiable by any configured instance type"), availableTypes, } - return } - 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) + sort.Slice(types, func(i, j int) bool { + if types[i].Price != types[j].Price { + // prefer lower price + return types[i].Price < types[j].Price + } + if types[i].RAM != types[j].RAM { + // if same price, prefer more RAM + return types[i].RAM > types[j].RAM + } + if types[i].VCPUs != types[j].VCPUs { + // if same price and RAM, prefer more VCPUs + return types[i].VCPUs > types[j].VCPUs + } + if types[i].Scratch != types[j].Scratch { + // if same price and RAM and VCPUs, prefer more scratch + return types[i].Scratch > types[j].Scratch + } + // no preference, just sort the same way each time + return types[i].Name < types[j].Name + }) + // Truncate types at maxPrice. We rejected it.Price>maxPrice + // in the loop above, but at that point maxPrice wasn't + // necessarily the final (lowest) maxPrice. + for i, it := range types { + if i > 0 && it.Price > maxPrice { + types = types[:i] + break + } } + return types, nil }