Network map
Federation network map
Every institution on a flat world map or a rotating globe, with the experiments each one runs, the kind of data it holds, the sample counts it declared, and whether its endpoint is answering. The coordinator and every partner open the same page.
The viewer
What the viewer shows
Globe and flat map
The same federation in two projections. The globe conveys the distances involved; the flat map is easier for reading labels and selecting sites. Arcs run from the coordinator to each participating site.
Experiments
Every site is a card: institution, city and country, the experiments it takes part in, and the samples it declared for each. Filtering by experiment redraws the map to that experiment's federation.
Results
Selected figures, tables and reports from completed runs, opened beside the map rather than distributed as a directory of files.
Runs
A live run streams onto the same map as it proceeds, per round and per site. A completed run can be replayed from its recorded events.
Design
Design properties
The map exists before any run does
It is drawn from federation.yaml alone, so a partner appears on the
map the day the configuration names them. This makes it a view of the network
rather than a dashboard for a single job: partners are recruited and stand up
endpoints well before anything is launched, and the map is populated throughout
that period.
Geography is declared, not detected
Site coordinates come from federation.yaml and are never resolved from
an IP address. A partner's compute node has no outbound web access, so the lookup
would fail; if it succeeded it would return the institution's border router rather
than the cluster; and it would add an outbound call from inside a partner's
security boundary for the sole purpose of drawing a dot. A site with no
coordinates is reported as unplaced rather than drawn at a guess.
What is deliberately not published
Site names, countries, coordinates, sample counts and experiment names are in the published data by design — that is the map. Service accounts, cluster paths, the coordinator's Globus identity and, by default, full endpoint UUIDs are not. Endpoint UUIDs are truncated to a fingerprint unless you opt in.
Export
Exporting your own
The export consists of three static files: the viewer, the metadata and an index. It requires no server and no Python at the other end. Placed on GitHub Pages, an institutional web host or any object store, it gives everyone in the federation the same page.
pip install "appfl-bio-suite[watch]==0.3.2" -c constraints.txt
appfl-bio-suite watch build # draw it from federation.yaml
appfl-bio-suite watch serve # look at it locally
appfl-bio-suite watch export --out site/ # three static files anyone can host
appfl-bio-suite run gwas --watch # a live run, on the same map
watch export writes out
the site names, countries, coordinates, sample counts and experiment names that
constitute the map. Pass --experiment to publish a single experiment's
federation, and read
docs/coordinator/network-map.md
before distributing the URL.
Appending ?view=globe to the viewer URL opens it on the globe, and
?theme=light selects the light palette. This page's embed and its full-screen
links use both, so the viewer opens in this page's theme.
Upstream
hivewatch and APPFL
HiveWatch
APPFL's monitoring toolkit, which supplies the event schema, the map metadata
format, the SSE server and the Leaflet viewer. This suite adds the federation
itself: hivewatch knows how to draw a client, and federation.yaml
records who the clients are.
APPFL
The federated learning framework underlying every experiment in the suite: the aggregators, the client trainers, and the communication layer this project drives over Globus Compute.
github.com/APPFL/APPFL ↗