The front decides.
lispers.net is the reference implementation of the LISP map-server and map-resolver. Ribbit-LISP does the same job against the same wire. Its state lives in shared memory on a separate host, and the front that receives LISP packets keeps no copy of it and takes no locks.
Measured over the Internet.
Load from New York; both map-servers on Seattle7 in Seattle; Ribbit’s memory on a two-core droplet in San Francisco. Every operation pays about 100 ms of network, and the closed-loop load caps any server near 1,500 operations a second on this path.
| lispers.net 0.643 | Ribbit-LISP v0.75 | |
|---|---|---|
| 78 behaviour tests | 14 failures; stopped serving 7 times | 0 failures; 53 byte-for-byte parity |
| lookups/s at peak | 87 (64 clients) | 1,468 (1,024 clients): 17× |
| registrations/s at peak | 186 (64 clients) | 1,474 (1,024 clients): 7.9× |
| 256 clients and up | nothing answered within 1 s | 1,357–1,479 operations/s |
| CPU per million lookups | 13,088 s | 96 s front + 4 s RAM host: about 140× less |
| 10,000 registrations | 6 of 25 sampled never resolved | all 25 resolved 2.9 s after the preload |
| application code (matched audit) | 6,463 NLOC, CCN 2,237 | 2,067 NLOC, CCN 1,318: 68% less code, 41% less complexity |
The earlier run: one host.
A Raspberry Pi 5 load generator, both systems on one host, Ribbit’s memory on the droplet. A different arrangement from the November run; the two tables are not comparable row for row.
| lispers.net 0.643 | Ribbit-LISP v0.63 | |
|---|---|---|
| 78 behaviour tests | 14 failures | 0 failures; 53 byte-for-byte parity |
| lookups/s, 64 clients (p50) | 484 (130 ms) | 20,407 (2.9 ms) |
| lookups/s at 10,000 registrations, 16 clients | 93 | 9,123 |
| CPU per lookup, 1 → 10,000 registrations | 1.6 → 11.4 ms | 44 → 47 µs |
| registrations/s, 1 / 64 clients | 357 / 718 | 39 / 1,565 |
The contract is the wire.
Kept
RFC 9300 and 9301 and the LCAF, DDT and NAT-traversal pieces lispers.net implements: Map-Register with authentication, Map-Request, Map-Notify, Info-Request, Map-Referral, sites, policies, named locators. lispers.net’s own behaviour is the oracle: 78 differential tests.
Gone
The private copies of registration state, the locks around it, the IPC between lispers.net’s processes and the timers that repair them. Sites, registrations, keys and policies are cells in the memory; each has one owner.
Decided at the front
In v0.75 the front decides each Map-Register where it arrives and writes back without waiting. Keys are configuration rows. The RAM host stores; it does not compute. That arrangement is the step 5e rewrite for LISP.
How it was measured.
The full report covers what was built, how both were tested, and how far the results can be read. The simulator runs Ribbit-LISP against lispers.net over central, edge and slow-edge links on every gate.
opt/frognet_semantic/ribbit/examples/lispers.net/RIBBIT-LISP-REPORT.md | the report: both runs, every table, section 9 on reading them |
opt/frognet_semantic/ribbit/examples/lispers.net/evidence/2026-11-14 | the 14 November evidence: results, acceptance run, audit, per-function CSVs |
opt/frognet_semantic/simulation/lisp_slow_tier.py | the simulator tier |
Bring your system, its contract and its tests.
I apply the method on this site to your system and tell you what a conversion would take. You get a written assessment containing:
- The contract: the API, protocol or behaviour that stays exactly as it is.
- The truth table: every fact the system keeps, who owns it, and who reads it.
- The acceptance oracle: the tests the converted system must pass, your own first.
- What disappears: the servers, queues, retries, caches and copies that shared memory makes unnecessary.
- A time and cost estimate for the conversion itself.
It is the architecture review on the rates table, applied to a conversion. The method behind it is on Converting a system; the two worked examples are PyTorch and lispers.net.