Monitoring & Self-Management (SkyEye.KOM)
Deep-dive page: how a serverless P2P system observes itself and reconfigures — tree-based aggregation, capacity search, closed-loop management, and protection of the monitoring data itself.
Supersession note: the original design is SkyEye ICPADS 2008; the management loop is IEEE P2P 2009; the engineering/scaling study is Christian Groß's thesis 2009; the authoritative consolidated description is the dissertation chapter Graffi 2010; integrity protection is Disterhoft 2016; the integration into the application is CCNC 2011.
The problem
In a network of autonomous, unreliable nodes there is no central point where state could be observed and behaviour controlled — yet quality goals (e.g. "hop count between 7 and 10", "peer-load standard deviation ≤ 500 %") require coordinated self-configuration under churn. SkyEye.KOM answers this as a monitoring and management over-overlay: a tree built on top of any KBR/DHT, without special node roles and without being woven into the routing layer.
Tree construction (no maintenance of its own)
- The identifier space is recursively divided into domains; each domain's coordinator is determined deterministically from the domain key and the responsibility information of the underlying DHT (ICPADS 2008).
- A unified [0,1] identifier space is mapped back onto the overlay space, which makes SkyEye overlay-independent (FreePastry in LifeSocial/LibreSocial; also Pastry in simulations).
- Result: a tree of O(log N) depth maintained solely by protocol messages — no separate tree-keeping traffic (Dissertation 2010).
What flows up: aggregated statistics
Updates are pushed bottom-up; every node aggregates children's values into
count, sum, min, max, sum-of-squares → mean, variance, standard deviation for the whole
system. Update size stays independent of the number of samples
(P2P 2009,
Dissertation 2010).
Two payload classes:
- aggregatable statistics — averages, standard deviations, confidence intervals;
- non-aggregatable capacities — CPU, bandwidth, storage, online time, carried with a TTL (ICPADS 2008).
What flows down: capacity search and configuration
- Capacity-based peer search: "give me n peers with CPU ≥ x, upload ≥ y, free disk ≥ z" with AND / OR / less-than / greater-than constraints; queries are answered on their way up the tree. This is how application layers pick suitable peers for role assignment (relay, storage, streaming) — used by the load-balancing design (load balancing & capacity).
- Closed-loop management: the tree root derives a new configuration from the aggregated global view and distributes it downward. Oscillation protection: changes are made effective, execution is delayed, and action is taken only when the slope of the value history is small (P2P 2009).
- Load balancing inside the tree: weak nodes are relieved via thresholds and support peers; children below a threshold are merged (Groß 2009).
Engineering: OSGi integration
Groß's thesis built the framework as OSGi bundles communicating only through interfaces,
integrated into LifeSocial.KOM and the distributed Eclipse IDE PIPE
(Groß 2009);
the application-level view is in CCNC 2011.
In current code this is the com.p2pframework.datacollection package — see
monitoring component; the Tester consumes its metrics
(runs/<id>/metrics) as described in testing methodology.
Measured behaviour
| Setting | Result |
|---|---|
| Simulation, 1000/5000/10000 peers (2008) | query answers 0–4 hops avg; ~98.5 % of listed peers online; 1–3 update msgs/peer/interval |
| Testbed, up to 500 peers, 5 s interval (Groß 2009) | ~2 hops, ~3 KB/s per peer; churn to 50 % tolerated; compression cuts up to 70 % of traffic |
| 30–60 s intervals | ~200 B/s per peer |
| Scaling caveat (Groß 2009) | concatenation-based capacity aggregation grows toward upper tree levels → needs filters/incremental updates to scale |
Protecting the monitoring data (DOMiNo)
An aggregator is an attractive target: manipulated reports skew the global view and thus the self-configuration. The Convex-Hull-watchdog work (Disterhoft, LCN 2016):
- attacker taxonomy: parent/child/random manipulation, random-acknowledgement and root manipulation, free-riding; reporting patterns empty / upwards / downwards / do-nothing;
- evaluation: PeerfactSim.KOM, Pastry, 5000 nodes, no churn, 228 parameter combinations, branching factors 8/32/96, 5–20 % malicious nodes;
- without protection the relative error of the global view reaches ~998 %; DOMiNo limits it to ~9 % at a high manipulation factor, and to 2/6/18 % (instead of 11/24/53 %) at factor 2 with 5/10/20 % attackers; best config: branching factor 96, Z-score threshold 3.5;
- limits: purely passive — cannot detect tree loops or leaf nodes that simply stop reporting.
Open work item #55 (decentralized reputation for storage/relay nodes) builds directly on these metrics — see security roadmap blog.
Beyond monitoring (dissertation context)
The dissertation pairs SkyEye with P3R3O.KOM (reliable long-term resource reservation under churn via adaptive redundancy of peer assignment) — the management counterpart of monitoring, relevant to replication decisions today (Dissertation 2010).