/skymap_scanner

A distributed system that performs a likelihood scan of event directions for IceCube real-time alerts using CPU cluster(s) and queue-based message passing.

Primary LanguagePython

GitHub release (latest by date including pre-releases) Lines of code GitHub issues GitHub pull requests

Skymap Scanner v3

A distributed system that performs a likelihood scan of event directions for real-time alerts using inter-CPU queue-based message passing.

Skymap Scanner is the computational core of the SkyDriver orchestration service.

skymap_scanner is a python package containing two distinct applications meant to be deployed within containers (1 skymap_scanner.server, n skymap_scanner.clients), along with skymap_scanner.utils (utility functions) and skymap_scanner.recos (icetray reco-specific logic). Additional, package-independent, utility scripts are in resources/utils/.

Queue Types

The default queue type in the container is RabbitMQ, since v3.1.0. Build Dockerfile_pulsar for a pulsar container.

RabbitMQ

Env variables

# export SKYSCAN_BROKER_CLIENT=rabbitmq  # rabbitmq is the default so env var is not needed
export SKYSCAN_BROKER_ADDRESS=<hostname>/<vhost>
export SKYSCAN_BROKER_AUTH=<token>
export EWMS_PILOT_QUARANTINE_TIME=1200  # helps decrease condor blackhole nodes
export EWMS_PILOT_TASK_TIMEOUT=1200

Currently, RabbitMQ uses URL parameters for the hostname, virtual host, and port ([https://]HOST[:PORT][/VIRTUAL_HOST]). The heartbeat is configured by EWMS_PILOT_TASK_TIMEOUT. This may change in future updates.

Python install:

pip install .[rabbitmq]

Pulsar

Env variables

export SKYSCAN_BROKER_CLIENT=pulsar
export SKYSCAN_BROKER_ADDRESS=<ip address>
export SKYSCAN_BROKER_AUTH=<token>
export EWMS_PILOT_QUARANTINE_TIME=1200  # helps decrease condor blackhole nodes
export EWMS_PILOT_TASK_TIMEOUT=1200

Python install:

pip install .[pulsar]

Example

This example is for the rabbitmq (default) broker. Steps for using a pulsar broker are similar and differences are noted throughout this example. The predominant difference is noted in Queue Types (Dockerfile_pulsar).

Example Startup

You will need to get a rabbitmq broker address and authentication token to pass to both the server and client. Send a poke on slack #skymap-scanner to get those!

1. Launch the Server

The server can be launched from anywhere with a stable network connection. You can run it from the cobalts for example.

Figure Your Args
Environment Variables
export SKYSCAN_BROKER_ADDRESS=BROKER_ADDRESS
# export SKYSCAN_BROKER_CLIENT=rabbitmq  # rabbitmq is the default so env var is not needed
export SKYSCAN_BROKER_AUTH=$(cat ~/skyscan-broker.token)  # obfuscated for security
export EWMS_PILOT_TASK_TIMEOUT=1200
Command-Line Arguments
    --client-startup-json PATH_TO_CLIENT_STARTUP_JSON \
    --cache-dir `pwd`/server_cache \
    --output-dir `pwd` \
    --reco-algo millipede_original \
    --event-file `pwd`/run00136662-evt000035405932-BRONZE.pkl  # could also be a .json file

NOTE: The --*dir arguments can all be the same if you'd like. Relative paths are also fine. NOTE: There are more CL arguments not shown. They have defaults.

client-startup.json

The server will create a PATH_TO_CLIENT_STARTUP_JSON file that has necessary info to launch a client. the parent directory of --client-startup-json needs to be somewhere accessible by your client launch script, whether that's via condor or manually.

Run It
with Singularity
singularity run /cvmfs/icecube.opensciencegrid.org/containers/realtime/skymap_scanner:x.y.z" \
    python -m skymap_scanner.server \
    YOUR_ARGS
or with Docker
# side note: you may want to first set environment variables, see below
./resources/launch_scripts/docker/launch_server.sh \
    YOUR_ARGS

NOTE: By default the launch script will pull, build, and run the latest image from Docker Hub. You can optionally set environment variables to configure how to find a particular tag. For example:

export SKYSCAN_DOCKER_IMAGE_TAG='x.y.z'  # defaults to 'latest'
export SKYSCAN_DOCKER_PULL_ALWAYS=0  # defaults to 1 which maps to '--pull=always'
export EWMS_PILOT_TASK_TIMEOUT=1200

2. Launch Each Client

The client jobs can submitted via HTCondor from sub-2. Running the script below should create a condor submit file requesting the number of workers specified. You'll need to give it the same SKYSCAN_BROKER_ADDRESS and BROKER_AUTH as the server, and the path to the client-startup json file created by the server.

Figure Your Args
Environment Variables
export SKYSCAN_BROKER_ADDRESS=BROKER_ADDRESS
# export SKYSCAN_BROKER_CLIENT=rabbitmq  # rabbitmq is the default so env var is not needed
export SKYSCAN_BROKER_AUTH=$(cat ~/skyscan-broker.token)  # obfuscated for security
export EWMS_PILOT_QUARANTINE_TIME=1200  # helps decrease condor blackhole nodes
export EWMS_PILOT_TASK_TIMEOUT=1200
Command-Line Arguments

See notes about --client-startup-json below. See client.py for additional optional args.

Run It
with Condor (via Singularity)

You'll want to put your skymap_scanner.client args in a JSON file, then pass that to the helper script.

echo my_client_args.json  # just an example
./resources/client_starter.py \
    --jobs #### \
    --memory #GB \
    --singularity-image URL_OR_PATH_TO_SINGULARITY_IMAGE \
    --client-startup-json PATH_TO_CLIENT_STARTUP_JSON \
    --client-args-json my_client_args.json

NOTE: client_starter.py will wait until --client-startup-json PATH_TO_CLIENT_STARTUP_JSON exists, since it needs to file-transfer it to the worker node. Similarly, the client's --client-startup-json is auto-set by the script and thus, is disallowed from being in the --client-args arguments.

or Manually (Docker)
# side note: you may want to first set environment variables, see below
./resources/launch_scripts/wait_for_file.sh PATH_TO_CLIENT_STARTUP_JSON 600
./resources/launch_scripts/docker/launch_client.sh \
    --client-startup-json PATH_TO_CLIENT_STARTUP_JSON \
    YOUR_ARGS

NOTE: By default the launch script will pull, build, and run the latest image from Docker Hub. You can optionally set environment variables to configure how to find a particular tag. For example:

export SKYSCAN_DOCKER_IMAGE_TAG='x.y.z'  # defaults to 'latest'
export SKYSCAN_DOCKER_PULL_ALWAYS=0  # defaults to 1 which maps to '--pull=always'

3. Results

When the server is finished processing reconstructions, it will write a single .npz file to --output-dir. See skymap_scanner.utils.scan_result for more detail.

4. Cleanup & Error Handling

The server will exit on its own once it has received and processed all the reconstructions. The server will write a directory, like run00127907.evt000020178442.HESE/, to --cache-dir. The clients will exit according to their receiving-queue's timeout value (SKYSCAN_MQ_TIMEOUT_TO_CLIENTS), unless they are killed manually (condor_rm).

All will exit on fatal errors (for clients, use HTCondor to manage re-launching). The in-progress pixel reconstruction is abandoned when a client fails, so there is no concern for duplicate reconstructions at the server. The pre-reconstructed pixel will be re-queued to be delivered to a different client.

In-Production Usage Note: Converting i3 to json and scaling up

You may want to run on events stored in i3 files. To convert those into a json format readable by the scanner, you can do

cd resources/utils
python i3_to_json.py --basegcd /data/user/followup/baseline_gcds/baseline_gcd_136897.i3 EVENT_GCD.i3 EVENT_FILE.i3

This will pull all the events in the i3 file into run*.evt*.json which can be passed as an argument to the server.

For now, it's easy to scale up using the command line. Multiple server instances can be run simultaneously and a separate submit file created for each one. To run N servers in parallel

export SKYSCAN_BROKER_ADDRESS=BROKER_ADDRESS
export SKYSCAN_BROKER_CLIENT=rabbitmq
export SKYSCAN_BROKER_AUTH=$(cat ~/skyscan-broker.token)  # obfuscated for security
ls *.json | xargs -n1 -PN -I{} bash -c 'mkdir /path/to/json/{} && python -m skymap_scanner.server --client-startup-json /path/to/json/{}/client-startup.json --cache-dir /path/to/cache --output-dir /path/to/out --reco-algo RECO_ALGO --event-file /path/to/data/{}'

Then, from sub-2 run ls *.json |xargs -I{} bash -c 'sed "s/UID/{}/g" ../condor > /scratch/$USER/{}.condor' using the template condor submit file below. Then you should be able to just run:

ls /scratch/$USER/run*.condor | head -nN | xargs -I{} condor_submit {}

executable = /bin/sh 
arguments = /usr/local/icetray/env-shell.sh python -m skymap_scanner.client --client-startup-json ./client-startup.json
+SingularityImage = "/cvmfs/icecube.opensciencegrid.org/containers/realtime/skymap_scanner:x.y.z"
environment = "SKYSCAN_BROKER_AUTH=AUTHTOKEN SKYSCAN_BROKER_ADDRESS=BROKER_ADDRESS EWMS_PILOT_TASK_TIMEOUT=1200 EWMS_PILOT_QUARANTINE_TIME=1200"
Requirements = HAS_CVMFS_icecube_opensciencegrid_org && has_avx
output = /scratch/$USER/UID.out
error = /scratch/$USER/UID.err
log = /scratch/$USER/UID.log
+FileSystemDomain = "blah"
should_transfer_files = YES
transfer_input_files = /path/to/json/UID/client-startup.json
request_cpus = 1
request_memory = 8GB
notification = Error
queue 300

You may also need to add this line to the condor submit file if running millipede_wilks as some resources have been removed from the image.

environment = "APPTAINERENV_I3_DATA=/cvmfs/icecube.opensciencegrid.org/data SINGULARITYENV_I3_DATA=/cvmfs/icecube.opensciencegrid.org/data I3_DATA=/cvmfs/icecube.opensciencegrid.org/data I3_TESTDATA=/cvmfs/icecube.opensciencegrid.org/data/i3-test-data-svn/trunk"

The extra envs for I3_DATA are to ensure it gets passed through for use inside the container. Additionally, ftp-v1 ice was introduced in v3.4.0.

Additional Configuration

Environment Variables

When the server and client(s) are launched within Docker containers, all environment variables must start with SKYSCAN_ in order to be auto-copied forward by the launch scripts. EWMS_-prefixed variables are also forwarded. See skymap_scanner.config.ENV for more detail.

Timeouts

The Skymap Scanner is designed to have realistic timeouts for HTCondor. That said, there are three main timeouts which can be altered:

    # seconds -- how long client waits between receiving pixels before thinking event scan is 100% done
    #  - set to `max(reco duration) + max(subsequent iteration startup time)`
    #  - think about starved clients
    #  - normal expiration scenario: the scan is done, no more pixels to scan (alternative: manually kill client process)
    SKYSCAN_MQ_TIMEOUT_TO_CLIENTS: int = 60 * 30  # 30 mins
    #
    # seconds -- how long server waits before thinking all clients are dead
    #  - set to duration of first reco + client launch (condor)
    #  - important if clients launch *AFTER* server
    #  - normal expiration scenario: all clients died (bad condor submit file), otherwise never (server knows when all recos are done)
    SKYSCAN_MQ_TIMEOUT_FROM_CLIENTS: int = 3 * 24 * 60 * 60  # 3 days
    #
    # seconds -- how long client waits before first message (set to duration of server startup)
    #  - important if clients launch *BEFORE* server
    #  - normal expiration scenario: server died (ex: tried to read corrupted event file), otherwise never
    SKYSCAN_MQ_CLIENT_TIMEOUT_WAIT_FOR_FIRST_MESSAGE: int = 60 * 60  # 60 mins

Relatedly, the environment variable EWMS_PILOT_TASK_TIMEOUT & EWMS_PILOT_QUARANTINE_TIME can also be configured (see 1. Launch the Server and 2. Launch Each Client).

Command-Line Arguments

There are more command-line arguments than those shown in Example Startup. See skymap_scanner.server.start_scan.main() and skymap_scanner.client.client.main() for more detail.

Runtime-Configurable Reconstructions

Recos are registered by being placed in a dedicated module within the skymap_scanner.recos sub-package. Each module must contain a class of the same name (eg: skymap_scanner.recos.foo has skymap_scanner.recos.foo.Foo) that fully inherits from skymap_scanner.recos.RecoInterface. This includes implementing the static methods: traysegment() (for IceTray) and to_pixelreco() (for MQ). The reco-specific logic in the upstream/pixel-generation phase is defined in the same class by the prepare_frames() (pulse cleaning, vertex generation) and get_vertex_variations() (variations of the vertex positions to be used as additional seeds for each pixel). On the command line, choosing your reco is provided via --reco-algo (on the server).

Making Branch-Based Images for Production-like Testing

If you need to test your updates in a production-like environment at a scale that isn't provided by CI, then create a branch-based image. This image will be available on Docker Hub and CVMFS.

Steps:

  1. Go to Actions tab
  2. Go to docker & singularity/cvmfs releases workflow tab (on left column)
  3. Click Run workflow, select your branch, and click the Run workflow button
  4. Wait for the workflow steps to complete
    • You can check the workflow's progress by clicking the top-most entry (there will be a small delay after the previous step)
  5. Check https://hub.docker.com/r/icecube/skymap_scanner/tags and/or CVMFS (the filepath will be the bottom-most line of https://github.com/WIPACrepo/cvmfs-actions/blob/main/docker_images.txt)

Note

The resulting image is specific to the branch's most recent commit. To test subsequent updates, you will need to repeat this process.

Data Types

These are the important data types within the scanner. Since memory-reduction is a consideration, some are persisted longer than others.

Pixel-Like

There are 5 data types to represent a pixel-like thing in its various forms. In order of appearance:

1. (nside, pixel_id)-tuple

  • generated by pixels.py

2. I3Frame

  • generated by PixelsToReco
  • introduces position-variations (eg: milipede_original)
  • M I3Frame : 1 (nside, pixel_id)-tuple
  • sent to client(s), not persisted on the server
  • ~800 bytes

3. SentPixelVariation

  • used for tracking a single sent pixel variation
  • 1 SentPixelVariation : 1 I3Frame
  • persisted on the server in place of I3Frame
  • ~50 bytes

4. RecoPixelVariation

  • represents a pixel-variation reconstruction
  • sent from client to server, persisted on the server
  • 1 RecoPixelVariation : 1 SentPixelVariation
  • ~50 bytes

5. RecoPixelFinal

  • represents a final saved pixel post-reco (on the server only)
  • 1 RecoPixelFinal : M RecoPixelVariation
  • These types are saved in nsides_dict (NSidesDict)
  • ~50 bytes

Sky Map-Like

Unlike pixel-like data types, these types are meant to exist as singular instances within the scanner.

nsides_dict (NSidesDict)

  • a dict of dicts containing RecoPixelFinal objects, keyed by nside & pixel id
  • exists on the server only
  • grows as the scan progresses
  • not persisted past the lifetime of a scan

skyreader.SkyScanResult

  • a class/object for using the result of a scan outside of the scanner (see icecube/skyreader)
  • created at the end of the scan (from nsides_dict)
    • intermediate/incomplete instances exist only to be sent to SkyDriver
  • can be exported to JSON and/or .npz-file
  • can be created from nsides_dict (internal to the scanner), JSON, and/or .npz-file
  • SkyDriver persists a serialized (JSON) version for each scan

Versioning

The MAJOR.MINOR.PATCH versioning scheme is updated according to the following

  1. MAJOR: Breaking change or other fundamental change in the skymap-scanner
  2. MINOR: Physics change or non-breaking new feature
  3. PATCH: Bug fixes

When the icetray image is updated, try to follow the same schema as its version update. So if icetray is bumped up a minor release, also increment it here.