import (
"errors"
- "log"
- "os/exec"
+ "regexp"
"sort"
- "strings"
- "time"
+ "strconv"
- "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")
AvailableTypes []arvados.InstanceType
}
+var pdhRegexp = regexp.MustCompile(`^[0-9a-f]{32}\+(\d+)$`)
+
+// estimateDockerImageSize estimates how much disk space will be used
+// by a Docker image, given the PDH of a collection containing a
+// Docker image that was created by "arv-keepdocker". Returns
+// estimated number of bytes of disk space that should be reserved.
+func estimateDockerImageSize(collectionPDH string) int64 {
+ m := pdhRegexp.FindStringSubmatch(collectionPDH)
+ if m == nil {
+ return 0
+ }
+ n, err := strconv.ParseInt(m[1], 10, 64)
+ if err != nil || n < 122 {
+ return 0
+ }
+ // To avoid having to fetch the collection, take advantage of
+ // the fact that the manifest storing a container image
+ // uploaded by arv-keepdocker has a predictable format, which
+ // allows us to estimate the size of the image based on just
+ // the size of the manifest.
+ //
+ // Use the following heuristic:
+ // - 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
+ return ((n - 80) / 42) * (64 * 1024 * 1024)
+}
+
+// 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 plus arv-mount block cache.
+func EstimateScratchSpace(ctr *arvados.Container) (needScratch int64) {
+ for _, m := range ctr.Mounts {
+ if m.Kind == "tmp" {
+ needScratch += m.Capacity
+ }
+ }
+
+ // Account for disk space usage by Docker, assumes the following behavior:
+ // - Layer tarballs are buffered to disk during "docker load".
+ // - Individual layer tarballs are extracted from buffered
+ // copy to the filesystem
+ dockerImageSize := estimateDockerImageSize(ctr.ContainerImage)
+
+ // The buffer is only needed during image load, so make sure
+ // the baseline scratch space at least covers dockerImageSize,
+ // and assume it will be released to the job afterwards.
+ if needScratch < dockerImageSize {
+ needScratch = dockerImageSize
+ }
+
+ // Now reserve space for the extracted image on disk.
+ needScratch += dockerImageSize
+
+ // Now reserve space the arv-mount disk cache
+ needScratch += ctr.RuntimeConstraints.KeepCacheDisk
+
+ return
+}
+
+// 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 cheapest available
// arvados.InstanceType big enough to run ctr.
func ChooseInstanceType(cc *arvados.Cluster, ctr *arvados.Container) (best arvados.InstanceType, err error) {
return
}
- needScratch := int64(0)
- for _, m := range ctr.Mounts {
- if m.Kind == "tmp" {
- needScratch += m.Capacity
- }
- }
+ needScratch := EstimateScratchSpace(ctr)
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
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 ok && it.Price > best.Price: // already selected a node, and this one is more 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.Price == best.Price && (it.RAM < best.RAM || it.VCPUs < best.VCPUs): // same price, worse specs
+ 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:
+ // Didn't reject the node, so select it
// Lower price || (same price && better specs)
best = it
ok = true
}
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)
- }
-}
projects / arvados.git / blobdiff