GitLab

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)

See Issue #10

`ContainerManager` Class

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

See Issue #10

Small fixes to correct for wrong actions on power control over a real
load.

This patch adds a poweractuator based on rapl settings available through
the sensor manager. Adding this actuator forces us to use a list of
actuators in the controller, changing a bit the structure of the code.

The "control" part of the NRM is bound to change and become more complex
in the near future, so move it in its own module.

This refactor also introduce some controller logic. Control is split
into 3 steps: planning, execution and updates. The goal is to use this
new code organization as a way to abstract different control policies
that could be implemented later.

Note that we might at some point move into a "control manager" and a
bunch of "policies" and "actuators", as a way of matching typical
control theory vocabulary.

Fixes a copy/paste mistake on the name of the callback to trigger on
stderr events.

This patch moves the tracking of applications clients of the downstream
API into a ApplicationManager, that is able to track progress and thread
management.

This change is necessary in the long term to build a comprehensive
downstream API and centralize the management of application tracking.

Note that this tracking is currently independent of the container and
pid tracking, and that might be a problem in the long term.

This patch refactors the downstream API to use pub/sub socket pair, like
the upstream API. This is part of the effort to improve the downstream
API. See #2.

This patch doesn't touch the client module, which will be adapted in
future commits.

The way 0MQ works on PUB/SUB sockets, publishers might drop
messages if subscribers are not detected faster enough. One way to fix
it is to have the "server" always bind sockets, and the "client" use
connect. This way, the handshake is initiated properly, and the client
can publish as soon as the connection is done.

This patch makes the daemon bind on the upstream API and the CLI connect,
fixing in the process the message dropping we were experiencing before.

Long term, we might have a think of using 2 types of sockets for the
upstream API: pub/sub for actual events published from the daemon, and
a REQ/REP or ROUTER/DEALER pair for "commands".

This patch implements a small finite state machine on the cmd side to be
able to run a command, wait for all of its output, and then exit.

As the daemon can send those message in any order, we need to wait them
properly, in particular the closing of stdout/stderr before exiting.

This patch also fixes the read_until_close callback creation to ensure
that the stream EOF is handled as a distinct message.

The daemon code was maintaining its own container tracker using pids,
instead of using the one in the container manager. This patch removes
this additional tracking, and let the daemon side deal with an actual
namedtuple.

This patch adds stdout/stderr streaming capabilities, based on partial
evaluation of a tornado.iostream callback. The bin/cmd CLI is updated to
wait until an exit message, although that doesn't guaranty anything on
message ordering...

The next step is obviously to figure out a message flow that allows the
CLI to send and receive the command IO properly, in order...

This patch propagates the process object into the container namedtuple,
fix a couple of bad function calls and adapt the run command handler to
use that process object instead of just the pid of it.

The logging improvement patch missed a few calls.

The logging module allow us to configure logging facilities once per
process using basicConfig, and then to use globally defined, named,
logger objects. This simplifies access to logger objects, their
configuration and remove pointers from all objects.

This patch refactor all the logging calls to use a single 'nrm' logger
object, using those facilities.

This patch adds a command to kill the parent process of a container
based on the container uuid, triggering the death of the container.

The os.kill command interacts pretty badly with the custom built
children handling, causing us to catch unwanted exceptions in an effort
to keep the code running. The waitpid code was also missing a bit about
catching children exiting because of signals, so we fixed that.

At this point, two things should be paid attention to:
  - we don't distinguish properly between a container and a command.
  This will probably cause issues later, as it should be possible to
  launch multiple programs in the same container, and for partitions to
  survive the death of the parent process.
  - the message format is growing more complex, but without any
  component having strong ownership over it. This will probably cause
  stability issues in the long term, as the format complexifies and we
  lose track of the fields expected from everyone.

This patch adds a very simple command to list the containers currently
known by the NRM. There's no history or state tracking on the NRM, so
the code is pretty simple.

We expect that some of the container tracking doesn't need to be sent
for such a command, so the listing also filters some of the fields.

This patch also adds an 'event' field to container messages, as it would
probably be needed further for other kind of operations.

This patch refactor the resource management and hwloc code into a
working, albeit very simple scheduling policy. Indeed, the previous code
contained strong assumptions about the output of hwloc matching an Argo
NodeOS configuration used during the previous phase of the project, that
always contained enough CPUs and Mems to perform exclusive scheduling.

The current version is simpler, but should work on more regular systems.
The patch also improves code organization so that introducing more
complex scheduling algorithms will be simpler.

The testing of this code resulted in the discovery of simple bugs in the
daemon children handling code, which should work now.

The Argus (globalos) launcher had prototype code to read a container
manifest, create a container using Judi's code, and map resources using
hwloc.

This patch brings that code, almost intact, into the NRM repo. This code
is quite ugly, and the resource mapping crashes if the kernel
configuration isn't right. But it's still a good starting point, and we
should be able to improve things little by little.

One part in particular needs attention: SIGCHLD handling. We should
think of using ioloop-provided facilities to avoid this mess.

The patch also contains the associated CLI changes.

Note: the messaging format is starting to be difficult to keep in check,
as there's conversions and field checks all over the code. See #3 for
a possible solution.

This is the first step in a series of patches to integrate the container
launching code from Argus (globalos) into the NRM infrastructure.

This patch creates a valid command on the CLI, and sends the necessary
info to the NRM. We still need to take care of the actual container
creation.

Note that the CLI waits for an event indicating that the container was
launched, at that at this point the event is never generated by the NRM.

This commit changes the message format for the upstream API, to use a
json-encoded dictionary. While the format is not set in stone at this
point, the goal is to slowly move into a proper protocol, with
well-defined fields to the messages, and proper mechanisms to send
commands and receive notification of their completion.

The only current user of this API is the power management piece, and
this change breaks the GRM code maintained outside of this repo. We will
need to reconcile the two implementation once the message protocol gets
more stable.

Related to #1 and #6.

Only supports setpower for now, and while it should work in theory, the
current code doesn't have a way to check if the command was received, as
the daemon never advertise the current limit.

We need to change the protocol at this point.

This also fixes a bug in the daemon code, that was expecting a single
string as a message, instead of a list of parts, as zmqstream always
receives.

Remove unneeded module, clean up calls to the logger.

The previous commit added pub/sub communications in the wrong places,
creating synchronizations in an asynchronous event loop.

This commit fixes those issues, adding the upstream (GRM/Flux) flow to
the event loop, and renaming objects here and there for clarity.

This patch is the last link between coolr and the daemon. We now create
a dictionary of machine information inside the sensor manager, and give
this dictionary back to the daemon. The daemon can then use the real
data for control.

We still need to receive the target power from somewhere, and that will
come later.

This a partial import of github.com/coolr-hpc/pycoolr from
master branch, commit id: 67e7aa4b89b67744922b5926cd1459adf650013b

Coolr will provide us the capability to read power, topology and
msr-based sensors.

This patch links the sensor module with the coolr code as it is, which
doesn't really work. The core issue is that the current coolr code is
meant to be stand-alone, and the main functions to both sampling and
formatting of the data in json, by hand.

We will solve that in the next commit, removing the json-specific code
from the sample function to create a dict of values instead.

Create a `sensor` module that will handle the interaction with coolr and
return updated machine information regularly.

The code code is as dumb as the previous one, but the structure is
improved, which should help for the next round of updates with coolr.

Giving the deadlines we have, it's probably a good idea to keep things
simple and keep the application facing protocol exactly as it is, and
restrict it to just applications. This way we can keep using the argobots
tests as valid benchmarks, and at the same start building a decent
communication protocol on a different socket, with a better interface.

To clarify that, the daemon now use the word application to refer to
clients connecting on the "legacy" interface.

We'll add a different socket and start building a real protocol in
future commits.

The previous code was entirely inside the bin directory, which is not a
good idea in the long term.

This patch move everything inside the nrm package, so that we can start
building a proper code.

We chose to rewrite the entire thing in python. The language should make
it easy to interact will all the moving parts of the Argo landscape, and
easy to prototype various control schemes.

The communication protocol is exactly the same, but implemented with
ZeroMQ + tornado.

Power readings are not integrated yet, we are targeting using the Coolr
project for that.

This is a rough draft, all the code is in binary scripts instead of
the package, and there are no unit tests. Nevertheless, it should be
a decent starting point for future development.