X-Git-Url: https://git.arvados.org/arvados.git/blobdiff_plain/df5c912a9eb5af7222e5446bc437ee97262542c8..6c78b28f9f54664babc57a4b4372c502065ed5d1:/lib/dispatchcloud/node_size.go diff --git a/lib/dispatchcloud/node_size.go b/lib/dispatchcloud/node_size.go index e77c862b36..6fb46b5f46 100644 --- a/lib/dispatchcloud/node_size.go +++ b/lib/dispatchcloud/node_size.go @@ -6,7 +6,9 @@ package dispatchcloud import ( "errors" + "regexp" "sort" + "strconv" "git.curoverse.com/arvados.git/sdk/go/arvados" ) @@ -22,6 +24,65 @@ type ConstraintsNotSatisfiableError struct { 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. +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 + + 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) { @@ -30,12 +91,7 @@ func ChooseInstanceType(cc *arvados.Cluster, ctr *arvados.Container) (best arvad return } - needScratch := int64(0) - for _, m := range ctr.Mounts { - if m.Kind == "tmp" { - needScratch += m.Capacity - } - } + needScratch := EstimateScratchSpace(ctr) needVCPUs := ctr.RuntimeConstraints.VCPUs