X-Git-Url: https://git.arvados.org/arvados.git/blobdiff_plain/a310d114bdc06b20cd007e6aff14b409e1c11e32..5ee9b22d0777d7ef42fa6acc05b666fd30e1ab2b:/lib/dispatchcloud/node_size.go diff --git a/lib/dispatchcloud/node_size.go b/lib/dispatchcloud/node_size.go index 6fb46b5f46..8cc63dc208 100644 --- a/lib/dispatchcloud/node_size.go +++ b/lib/dispatchcloud/node_size.go @@ -10,7 +10,7 @@ import ( "sort" "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") @@ -83,6 +83,19 @@ func EstimateScratchSpace(ctr *arvados.Container) (needScratch int64) { 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) { @@ -96,19 +109,34 @@ 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 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