GitLab

Resolve "Bad relative path handling between command line client and daemon"

Closes #8

See merge request !38

This is not as good as passing part of the manifest options forward,
but it still fixes some of the practical problems when using the
components together. The code makes sure that the manifests exists,
though.

[feature] commandline arguments, config file management.

See merge request !37

Added integration tests to the `.gitlab-ci.yaml` file. These tests
run any runner tagged "integration". They depend on the 'argotest'
repository master branch being able to run the tests, which
uses 'argopkgs' repository. The runner must be Nix
enabled. This therefore currently uses a rolling release of the
integration tests.

`cmd` now sends a container kill message to the upstream api and exits
whenever it receives SIGINT, via C-c for instance.

Fixing a bug introduced by the 'progress-report' branch in a recent
previous commit. The process object is the result of a tornado spawn, so
the call has to be slightly different than what was there.

This commit adds a command-line interface to `daemon`:
```
usage: daemon [-h] [-c FILE] [-d] [--nrm_log NRM_LOG] [--hwloc HWLOC]
              [--argo_nodeos_config ARGO_NODEOS_CONFIG] [--perf PERF]
              [--argo_perf_wrapper ARGO_PERF_WRAPPER]

optional arguments:
  -h, --help            show this help message and exit
  -c FILE, --configuration FILE
                        Specify a config json-formatted config file to
                        override any of the available CLI options. If an
                        option is actually provided on the command-line, it
                        overrides its corresponding value from the
                        configuration file.
  -d, --print_defaults  Print the default configuration file.
  --nrm_log NRM_LOG     Main log file. Override default with the NRM_LOG.
                        environment variable
  --hwloc HWLOC         Path to the hwloc to use. This path can be relative
                        and makes uses of the $PATH if necessary. Override
                        default with the HWLOC environment variable.
  --argo_nodeos_config ARGO_NODEOS_CONFIG
                        Path to the argo_nodeos_config to use. This path can
                        be relative and makes uses of the $PATH if necessary.
                        Override default with the ARGO_NODEOS_CONFIG
                        environment variable.
  --perf PERF           Path to the linux perf tool to use. This path can be
                        relative and makes uses of the $PATH if necessary.
                        Override default with the PERF environment variable.
  --argo_perf_wrapper ARGO_PERF_WRAPPER
                        Path to the linux perf tool to use. This path can be
                        relative and makes uses of the $PATH if necessary.
                        Override default with the PERFWRAPPER environment
                        variable.
```

Improve process management

See merge request !30

- added correct SIGINT/process ending handling to cmd
- fixed  kill/list containers
- added ZMQ_LINGER 0 to the socket options.

Related to #22

Add a listen command to get access to the event stream of the upstream
pub/sub API.

This patch gives back access from the command line to the power
information of a container, including filtering the event stream to only
have events relevent to this container.

This changes the workflow a little bit for users, but should result in a
cleaner access to profiling data in the future.

Related to #18.

Make it so that the daemon will delete containers when all commands it
is aware of are finished, instead of relying on a single owner that
needs to be tracked.

This simplifies the handling to multiple commands in the same container,
and should not impact the rest.

Move the container start/exit events to the upstream pub/sub event
stream. As these are more of a global event now that we support multiple
commands in the same container, it makes sense to move them to the more
general event stream.

This patch also remove the code in cmd waiting for container start or
exit, making (temporarily) the cmd unable to report power metrics. We
will fix that in a later commit.

This patch fixes complicated issues we had with how a job running
multiple commands in the container might not all wait for the end of the
container: now none of them do.

Add a upstream pub client, to be able to listen to messages coming from
the daemon on the upstream pub/sub channel.

Doesn't support any fancy filter, as that's not used by the daemon so
far.

Ensure that the client that created the container is considered as the
one owning it, with the consequence that if its command exits, the
container is destroyed. Also deals with the race issue we had on the cmd
side.

Current code sends start/exit events when a container is created and
process_start/process_exit when its already there. Instead, have the
container start/exit only care about container stuff, and always sends
the process start/exit events around. That makes the cmd run fsm easier
to work out.

Changes the message format a tiny bit.
Fixes some missing stdout/stderr issues we had before.

Previous merges let the cmd send an empty container uuid, resulting in
some issues when the user doesn't provide one. Restore the previous
behavior.

[Removing old code] removing old downstream API c code.

See merge request !31

[Fix] Handle container with no power profiling

See merge request !29

Handles container with no power profiling enabled in the manifest file.
In such cases the 'exit' response on process termination would generate
TypeError.

Improve Messaging layer

See merge request !28

Replace the fragile upstream communications with the new messaging
layer, improving the stability and performance of this API.

NOTE: this breaks previous clients
NOTE: this patch is missing client tracking, to handle children signals.

Abstracts away the exact wire format and client/server details, while
changing the RPC side to work over ROUTER/DEALER sockets, as to avoid
the lost messages issues we've been having with PUB/SUB for RPC.

[Feature] Multi- node and process support

Closes #17

See merge request !27

Added multi- node and process support that will allow launching of
multiple processes within a container. This is important for enabling
use of NRM with MPI applications with multiple processes in a container
and thus enabling multi-node executions

See Issue #17

[fix] Lock with Pipenv all packages

See merge request !25

Some packages were being pulled into the virtual environment through
setup.py. It just happens that the new versions of pyzmq and tornado
don't really play nice with the current daemon code. Mostly because of a
"better" ioloop hijacking mecanism, that doesn't work for us.

This patch moves the install requirements of the setup.py to Pipfile,
and lock them to a working version.

Note: the code still currently triggers #14.

Add application level power profiling support

Closes #12

See merge request !24

Added power profile data to the exit event response at the end of
application run. The profile data is generated using functions from
`SensorManager` and is obtained using the sensor update in `Daemon`

See Issue #12

Made changes in `SensorManager` to allow calculation of the difference in measured
(stored) values using `rapl_reader` functions

See Issue #12

Added profiling parameter to manifest file that can be used to specify
if power profile is required for a container/application. Also, changed
namedtuple "Container" key from "powerpolicy" to "power" and made
changes to reflect this

See Issue #12

[CI] ensure py.test is ran on runner with rapl

See merge request !22

Since our tests include checking that coolr can access the rapl
interface, we should ensure that only runners having rapl will test the
code.

Resolve "Integration of power policy in to NRM"

Closes #10

See merge request !20

Made changes in NRM to respond to phase_context event from [previously
called power_policy event (874a6a4d)] from the application. The NRM can now
store the informaton received on the event and call DDCM power policy through
interfaces developed (Issue #11) in the control loop

See Issue #10

Fixed the C downstream_api messages sent by the application to NRM to be more
generic. These messages can be used to invoke power or any other policies
relying on contextual information from an application phase

See Issue #10

Added initialization of power policy using the manifest file and fixed
appropriate TODOs in `PowerPolicyManager` (PPM) Class using container
information (except adding new power policies). Also, fixed PPM to use None
instead of NONE

See Issue #10

The container creation information (resources and powerpolicy) added to
`ContainerManager` class can be used to intiliaze power policy parameters
in the present and more in the future

See Issue #10