Skip to content
Permalink
Branch: master
Find file Copy path
Find file Copy path
Fetching contributors…
Cannot retrieve contributors at this time
236 lines (181 sloc) 8.15 KB

Note: This started out life a toy. Be aware that the code and the README won't always be in sync, following the instructions will likely expose something that needs to be tweaked or fixed before things work. See two related blog posts for more information:

This repository provides a compute workload scheduler using OpenStack placement to pick targets (via eschedule selecting and ecompute accepting), etcd as a transport, and the virt-install tool from virt-manager to run simple VMs.

It has been built to experiment with the idea of using placement and etcd as the main motors and state maintainers of a compute service and come to grips with some of the systems and process involved in creating virtual machines.

It assumes you've got a working libvirt install, with virt-install and virt-resize. On ubuntu that means the virtinst and libguestfs-tools packages. Note: If this ever gets to be slightly more than a toy using those command line tools should be changed to in-Python code.

Architecture

The overall architecture of the system is as follows. An etcd server is provided (via a docker container). A scheduler command line tool and multiple compute servers are clients of etcd, using watchers to notify the computes.

A placement service runs, also in a container, talking to a database.

One or more ecompute processes start up and register themselves as resource providers with some inventory (calculated via the python psutil package) and then watch for new data at keys associated with themselves within etcd. A pool of workers is also started. These are responsible for creating or destroying VMs while the main process is responsible for interacting with etcd. This makes it possible to concurrently launch multiple VMs. More importantly it means that concurent requests to launch are not lost (this was true in earlier versions).

The console-script eschedule accepts an input of resource requirements and an optional image URL, requests allocation candidates from placement, attempts to claim the first one and if successful puts a value to the etcd key associated with the target compute node. The value is the resource requirements, the instance uuid and an image reference.

ecompute notices the new value on the watched key, retrieves a copy of the image, launches a VM using virt-install, and sets a key back on etcd saying so, and recording the IP of the guest.

A simple metadata server runs to keep booting of cloud images fast, and if the image supports it, let cloud-init do things like set an authorized ssh key. This should be replaced.

Trying It

To try it out yourself you need docker and a database, the code within this repo, and the Python requirements listed in setup.py (python setup.py develop will get them). The docker containers provide etcd and placement.

The containers can be started by running docker.sh. Placement will be at http://localhost:8080/.

Edit mdserver.conf as required and start the metadata server with (this will be improved or replaced):

sudo ip addr add 169.254.169.254 dev virbr0
sudo python md_server/mdserver/server.py mdserver.conf &

Configure ecompute by creating a compute.yaml in the same directory, looking something like this:

etcd:
  host: ds1
placement:
  endpoint: http://ds1:8080

If you do not create the file, defaults will be used, pointing to localhost.

If you do not want to resize disk images (it saves time but makes the reporting of disk usage inaccurate) add the following to compute.yaml

resize: False

Start ecompute on one or more hosts. Each host must have the python requirements, the virt-install related tools, and a compute.yaml pointing to placement and etcd. You can install the python requirements and the ecompute and eschedule console scripts with python setup.py develop.

ecompute

Because ecompute inspects the system for a real inventory, you end up multiple-booking inventory if you have more than one ecompute on same host, but it is possible to do so for testing.

By default, each time an ecompute is started a new resource provider is created. This means that there can be orphaned providers in placement that will be scheduled to, but don't have any listeners. You can work around this by calling delete from resource_providers; in the database, as required.

Or (new feature!), if uuid in compute.yaml is set to a specific value, that value will be used each time ecompute is started. If a resource provider with that uuid already exists, it's inventory will be (re)set to whatever is accurate. When ecompute gets a SIGINT or Ctrl-C it will lock its inventory by setting the reserved value on the VCPU inventory to equal total and then exit. Next time it is started (with the same uuid), reserved will be cleared.

Once ecompute is running, we can try to schedule a workload. eschedule can run from any host that has network access to the placement and etcd servers. Modify schedule.yaml as required.

eschedule 'resources=VCPU:1,DISK_GB:1,MEMORY_MB:256'

You can also request an image:

eschedule resources=VCPU:1,MEMORY_MB:256,DISK_GB:1 \
    https://cloud-images.ubuntu.com/bionic/current/bionic-server-cloudimg-amd64.img

If no image is specified Cirros 0.3.6 is used.

The output from eschedule will look something like this:

NOTIFIED TARGET, b8756be5-a30d-4311-920c-0ad996367a8e, \
  OF INSTANCE d578fb7c-7787-4e73-b69a-a7b3ef9bf73a

And from ecompute:

INSTANTIATE INSTANCE d578fb7c-7787-4e73-b69a-a7b3ef9bf73a \
  WITH IMAGE e9bedb97-917a-4f4e-9a69-8ad21840267f
ALLOCATIONS ARE {'b8756be5-a30d-4311-920c-0ad996367a8e': \
  {'resources': {'VCPU': 1, 'DISK_GB': 1, 'MEMORY_MB': 256}}}

If there is no capacity available, either because there's none left or because the resource requirements are too strict it will look like this:

eschedule \
  'resources=VCPU:1,DISK_GB:1,MEMORY_MB:256&required=MISC_SHARES_VIA_AGGREGATE'
NO ALLOCATIONS LEFT

After an instance is booted its IP address is put back in etcd, which you can query by giving the instance uuid to eschedule:

eschedule d578fb7c-7787-4e73-b69a-a7b3ef9bf73a

By default the guest IP is only accessible from the host. If you define a bridge interface in compute.yaml this can be worked around. See BRIDGE.md for more.

You can destroy an instance by:

eschedule destroy d578fb7c-7787-4e73-b69a-a7b3ef9bf73a

This will destroy and undefine it on the host, remove the disk, and clear the allocations in placement. You can also use virsh to destroy VMs, but this will not clean up allocations.

Note: The database and etcd data (in /data/etcd) are not cleaned up. You'll want to take care of that yourself.

Things to Clean Up

  • On startup a compute should check to see if the metadata server is there, and if not, fork and start one.
  • Resizing disk images is currently done with multiple subprocess calls, this is cumbersome and weird.
  • Switch all the subprocess calls to using the python libvirt package directly.
  • Asking to destroy a VM while it is being built has not been tested and is likely to go poorly.

Concepts

  • Configuration should be limited, because that's hassle.
  • Features should be limited, because that's hassle.
  • Failure is just failure. Try again yourself, we're not going to do it for you.

Misc

  • If you are running a linux VM on an esxi hypervisor as a host for this stuff you might need to export LIBGUESTFS_BACKEND_SETTINGS=force_tcg to get filesystem manipulation to work well.

Help Wanted

If you think this is interesting, and would like to help out, please feel to make a pull request. There are a few main areas that need attention:

  • Making networking more useful.
  • Error handling.
  • Interacting with libvirt and avoiding subprocess calls.
  • Properly choose and launching images with user-data.
You can’t perform that action at this time.