💙💛
drmaa2os - A Go API for OS Processes, Docker Containers, Cloud Foundry Tasks, Kubernetes Jobs, Open Cluster Scheduler / Grid Engine Jobs, Podman containers, and more...
DRMAA2 for OS processes and more
Update: The Go DRMAA2 interface and the implementation based on the JobTracker interface are now decoupled. In order to use a specific backend, like Docker, the package providing the JobTracker implementation needs to be imported so that the init() method is called for registering at the DRMAA2 implementation.
Like when using the Docker backend:
_ "github.com/dgruber/drmaa2os/pkg/jobtracker/dockertracker"
This is a Go API based on an open standard (Open Grid Forum DRMAA2) for submitting and supervising workloads running in operating system processes, containers, PODs, tasks, or HPC batch jobs.
The API allows you to develop and run job workflows in OS processes and switch later to containers running in Kubernetes, as Cloud Foundry tasks, pure Docker, Singularity, or any HPC workload manager which supports the DRMAA standard through the C libdrmaa.so library (like SLURM, Open Cluster Scheduler, Grid Engine, ...) without changing the application logic.
Its main purpose is supporting you with an abstraction layer on top of platforms, workload managers, and HPC cluster schedulers, so that a software developer don't need to deal with the underlaying details and differences of job submission, status checking, and more.
An even simpler interface for creating job workflows without dealing with the DRMAA2 details is wfl which is based on the Go DRMAA2 implementation.
For details about the mapping of job operations please consult the platform specific READMEs:
In a very experimental state are:
- Split trackers in front/backend via OpenAPI
- New Containerd
- Singularity
- Podman
- mpioperator
- Archived Cloud Foundry
Note, that Singularity/Apptainer (or Podman) should be considered to be handled with the regular OS process backend, just starting the container cli. You gain features like job array task throttling and more...
Feedback welcome!
For a Go DRMAA2 wrapper based on C DRMAA2 (libdrmaa2.so) like for Univa Grid Engine please check out drmaa2.
Following example demonstrates how a job running as OS process can be executed. More examples can be found in the examples subdirectory. Per default jobs are managed in main memory hence after restarting your app all processes are not visible to your app even they are running. If persistency between restarts is required, please use NewDefaultSessionManagerWithParams() with a simpletracker.SimpleTrackerInitParams as argument.
Note that at this point in time only JobSessions are implemented.
import (
"github.com/dgruber/drmaa2os
_ "github.com/dgruber/drmaa2os/pkg/jobtracker/simpletracker"
)
sm, err := drmaa2os.NewDefaultSessionManager("testdb.db")
if err != nil {
panic(err)
}
js, err := sm.CreateJobSession("jobsession", "")
if err != nil {
panic(err)
}
jt := drmaa2interface.JobTemplate{
RemoteCommand: "sleep",
Args: []string{"2"},
}
job, err := js.RunJob(jt)
if err != nil {
panic(err)
}
job.WaitTerminated(drmaa2interface.InfiniteTime)
if job.GetState() == drmaa2interface.Done {
job2, _ := js.RunJob(jt)
job2.WaitTerminated(drmaa2interface.InfiniteTime)
} else {
fmt.Println("Failed to execute job1 successfully")
}
js.Close()
sm.DestroyJobSession("jobsession")Using other backends for workload management and execution only differs in creating a different SessionManager. Different JobTemplate attributes might be necessary when switching the implementation. If using a backend which supports container images it might be required to set the JobCategory to the container image name.
If Docker is installed locally it will automatically detect it. For pointing to a different host environment variables needs to be set before the SessionManager is created.
"Use DOCKER_HOST to set the url to the docker server. Use DOCKER_API_VERSION to set the version of the API to reach, leave empty for latest. Use DOCKER_CERT_PATH to load the TLS certificates from. Use DOCKER_TLS_VERIFY to enable or disable TLS verification, off by default."
import (
"github.com/dgruber/drmaa2os
_ "github.com/dgruber/drmaa2os/pkg/jobtracker/dockertracker"
)
sm, err := drmaa2os.NewDockerSessionManager("testdb.db")
if err != nil {
panic(err)
}
js, err := sm.CreateJobSession("jobsession", "")
if err != nil {
panic(err)
}
jt := drmaa2interface.JobTemplate{
RemoteCommand: "sleep",
Args: []string{"2"},
JobCategory: "busybox",
}
job, err := js.RunJob(jt)
if err != nil {
panic(err)
}
job.WaitTerminated(drmaa2interface.InfiniteTime)
js.Close()
sm.DestroyJobSession("jobsession") import (
"github.com/dgruber/drmaa2os
_ "github.com/dgruber/drmaa2os/pkg/jobtracker/kubernetestracker"
)
sm, err := drmaa2os.NewKubernetesSessionManager("testdb.db")
if err != nil {
panic(err)
}
js, err := sm.CreateJobSession("jobsession", "")
if err != nil {
panic(err)
}
jt := drmaa2interface.JobTemplate{
RemoteCommand: "sleep",
Args: []string{"2"},
JobCategory: "busybox",
}
job, err := js.RunJob(jt)
if err != nil {
panic(err)
}
job.WaitTerminated(drmaa2interface.InfiniteTime)
js.Close()
sm.DestroyJobSession("jobsession")The Cloud Foundry SessionManager requires details for connecting to the Cloud Foundry cloud controller API when being created. The JobCategory needs to be set to the application GUID which is the source of the container image of the task.
import (
"github.com/dgruber/drmaa2os
_ "github.com/dgruber/drmaa2os/pkg/jobtracker/cftracker"
)
sm, err := drmaa2os.NewCloudFoundrySessionManager("api.run.pivotal.io", "user", "password", "test.db")
if err != nil {
panic(err)
}
js, err := sm.CreateJobSession("jobsession", "")
if err != nil {
panic(err)
}
jt := drmaa2interface.JobTemplate{
RemoteCommand: "dbbackup.sh",
Args: []string{"location"},
JobCategory: "123CFAPPGUID",
}
job, err := js.RunJob(jt)
if err != nil {
panic(err)
}
job.WaitTerminated(drmaa2interface.InfiniteTime)
js.Close()
sm.DestroyJobSession("jobsession")Note, for Singularity it is recommended to use the process backend (simpletracker / calling Singularity command line interface) as the Singularity implementation was an experiment and is not tested with newer versions of Singularity.
The Singularity SessionManager wraps the singularity command which is required to be installed. The container images can be provided in any form (like pointing to file or shub) but are required to be set as JobCategory for each job.
import (
"github.com/dgruber/drmaa2os
_ "github.com/dgruber/drmaa2os/pkg/jobtracker/singularity"
)
sm, err := drmaa2os.NewSingularitySessionManager("testdb.db")
if err != nil {
panic(err)
}
js, err := sm.CreateJobSession("jobsession", "")
if err != nil {
panic(err)
}
jt := drmaa2interface.JobTemplate{
RemoteCommand: "sleep",
Args: []string{"2"},
JobCategory: "shub://GodloveD/lolcow",
}
job, err := js.RunJob(jt)
if err != nil {
panic(err)
}
job.WaitTerminated(drmaa2interface.InfiniteTime)
js.Close()
sm.DestroyJobSession("jobsession")The LibDRMAASessionManager can be used for submitting jobs through a pre-existing libdrmaa.so which is available and supported by many HPC workload managers (like Univa Grid Engine, SLURM, PBS, LSF, Son of Grid Engine, ...).
There are a few things to consider at compile time and runtime. The CGO_LDFLAGS and CGO_CFLAGS must be set according to the documentation in https://github.com/dgruber/drmaa. Also the LD_LIBRARY_PATH needs to be set accordingly.
An example using Grid Engine running in a container is here
The compile time configuration is external meaning the C library must be in the path or LD_LIBRARY_PATH and CGO_LDFLAGS and CGO_CFLAGS must be set according to the documentation in https://github.com/dgruber/drmaa.
import (
"github.com/dgruber/drmaa2os
_ "github.com/dgruber/drmaa2os/pkg/jobtracker/libdrmaa"
)
sm, err := drmaa2os.NewLibDRMAASessionManager("testdb.db")
if err != nil {
panic(err)
}First experimental version is implemented and tested on macos accessing Podman on a remote VM. When compiling on macos brew install gpgme helped me getting the C header dependencies of Podman installed. Accessing podman can be achieved through ssh in that case (calling podman system service --time=0 unix:///tmp/podman.sock in the podman VM for which the ssh port is defined at localhost:2222 on a Vagrant based vbox VM).
If ConnectionURIOverride is not set the implementation uses the default connection to the Podman REST API server. This server can be setup by podman system service -t 0 & in Linux environments.
Note, that it currently the implementation expects that the images are pre-pulled.
For running podman locally the process based implementation (simpletracker) can be used.
import (
"github.com/dgruber/drmaa2os
_ "github.com/dgruber/drmaa2os/pkg/jobtracker/podmantracker"
)
sm, err := drmaa2os.NewPodmanSessionManager(PodmanTrackerParams{
ConnectionURIOverride: "ssh://vagrant@localhost:2222/tmp/podman.sock?secure=False",
}, "testdb.db")
if err != nil {
panic(err)
}The remote directory in /pkg/jobtracker contains a client/server implementation of the JobTracker interface allowing to create clients and server for any backends (JobTracker implementations) mentioned above. The client/server protocol is defined in OpenAPI v3. Based on that Go client and server stubs have been generated using oapi-codegen. The OpenAPI spec contains also the DRMAA2 data types which might be useful for other projects.
The remote JobTracker server can be used in any Go DRMAA2 application.
import (
"github.com/dgruber/drmaa2os
_ "github.com/dgruber/drmaa2os/pkg/jobtracker/remote/client"
)
sm, err := drmaa2os.NewRemoteSessionManager(ClientTrackerParams{
Server: "localhost:8080",
}, "testdb.db")
if err != nil {
panic(err)
}The server can be implemented by using any JobTracker implementation as argument in the server implementation.
import(
"github.com/dgruber/drmaa2os/pkg/jobtracker/remote/server"
genserver "github.com/dgruber/drmaa2os/pkg/jobtracker/remote/server/generated"
)
func main() {
SetupHandler(simpletracker.New("jobsession"))
}
func SetupHandler(jobtracker jobtracker.JobTracker) {
impl, _ := server.NewJobTrackerImpl(jobtracker)
s := &http.Server{
Addr: ":8080",
Handler: genserver.Handler(impl),
ReadTimeout: 10 * time.Second,
WriteTimeout: 10 * time.Second,
MaxHeaderBytes: 1 << 20,
}
log.Fatal(s.ListenAndServe())
}