A fabric (NNI) port comes in 10G, 100G or 400G, as fast as the router
takes and no faster than 400G, and bundles into a LAG of up to eight
members like transit, IX and PNIs. Each member takes a router port (10G
on a breakout, a quarter of one) and pays the monthly price; upgrading
moves and pays for every member. 10G and 100G keep their prices, 400G is
$3,000/mo. A model change that would strand a 400G fabric port is refused,
and saves drop fabric ports that make no sense.
The NNI requirement counted what the port already carried but not the
offer's own bandwidth, so an offer passed as long as the port was not
already over and auto-accept took customers that did not fit. It now
counts the offer too.
Closes#140, closes#307.
From customer level 2, one person runs sales at a time, next to the NOC intern:
- The engineer turned salesdroid ($9k/mo) books you into every network operators' meeting 30 days
ahead at the usual ticket price, brings 40% more offers (half as many again while a meeting's glow
lasts) but never the huge deals, and their customers leave less often at renewal (half as often
for the nerdy ones: HFT, ISPs, universities, L2). Haggles badly.
- The shark ($15k/mo) brings bigger sizes at 25% more, takes a month of each deal as commission,
oversells half of them with a requirement the network does not meet yet, never brings technical
customers, makes selective peering desks want 25% more traffic, wins most haggles, and their
customers leave 50% more often.
- The CEO (free) signs the biggest internet customer the network can take on the golf course every
50-80 days at 30% off, often with a promise nobody checked, signs about a third of the offers you
decline anyway, and has opinions. He cannot be fired: a board meeting brings a VP of Sales a month
later.
Customers remember who signed them (by), which the renewals read. Every roll is a hash of the day
or the deal, so games without a salesperson keep their random stream.
The differential fuzz also leaves out buttons for actions main does not have yet.
What a scrubbing center sends back to its target is the target's own
traffic, but compute() carried it under the junk bucket, so a small
flood at a big customer showed up as hundreds of gigabits of DDoS junk
on the backbone. It now has its own bucket, shown as scrubbed clean
traffic in the link panel.
R.egT and R.netEg were summed by key in push(), which boxes every new value; they are now Float64Array sums written
back after routing, in the same order and key order. The first group G was emptied with length=0 for every flow,
which gives back its storage so the next push allocates it again; it is now G[0..GN). 27.5 MB -> 13.5 MB per compute.
compute() was one 33 KB function that V8 never kept optimised: it ran in baseline code, boxing every number,
and was recompiled in the background (about a second of compile each time on a loaded machine) only to deopt.
Its stages now run as closures of their own (compute's own bytecode: 7.7 KB) and the money is computeMoney().
The shortest paths and the exit lists (okBy, reach, BRJ/LIST, DSEL) are kept between calls under keys built
from everything they depend on (29-route-cache.js). Exits and paths have one object shape each.
The equivalence tests compare cached against uncached routing; tests/route-cache.test.js checks the keys.
The day is split around its routing (dayPre, compute, dayPost). In play, S goes to a
worker started from a Blob URL of the page's inline game script (nothing fetched), the
day waits for R and finishes when it arrives. Actions, failures, events and the intern
still route synchronously; an action mid-day wins over the worker's stale answer.
R crosses as plain data: R.ef as its raw arrays (transferred), flows as columns.
Tests and any worker failure keep the synchronous path.
Fewer short-lived objects per compute: no array growth while gathering flows
(copied into each result's own columns), no iterator results in push or in
the exits-by-network setup, and R.efg reads the edge sums in place. Results
are identical to main.
Flows are packed into Float64Array/Int32Array columns; R.fl builds the old
objects only when read. Edge egress sums keep a slot per egress in a plain
array, and R.tNet is added up in a typed array and written out at the end,
in the same key order. Results are identical to main.
Review of #205:
- Prefer no longer switches the carrier-direct rule off for the whole network.
A POP with a Prefer-marked transit port keeps what Prefer says; every other
POP keeps the rule. compute(), traceroute and the reference loop in the tests
all use the same per-POP rule, and a new test covers it (New York and Chicago,
Prefer on one POP's Tell-Ya port).
- dayTick() cost. The harness is neutral (the same page on both sides differs by
1%), but the cost was real: the four extra rank passes inside compute()'s main
loop made V8 deoptimise compute() 954 times in 30 game days against 46 on main.
The direct ranks now run in a small function of their own, in the plain way the
reference loop does (their lists are one carrier's ports), so the main loop is
the one main has, and the setup moved out of compute(). 61 deopts instead of 954.
- Italian: "Porta qui" read as "port here" next to "Porta più vicina"; it is
"Instrada qui" now, and "Ogni POP instrada il traffico..." matches.
- Help text: says that real BGP carries the traffic from every POP (the shorter
AS path beats IGP distance), that Carry here does that here, and that Nearest
port is the default. All six languages.
The merge kept main's version of compute(), the link actions and panels, the
traceroute exit list and the CHANGELOG, so the city panel called a carryCtl()
that no longer existed and routing never ranked a carrier's own port. Put the
ranking back into compute() on top of the current rank lists and island split,
along with setCarry, carryCtl, the traceroute carry flag, the reference loop in
the routing test and the CHANGELOG bullet. Drop a stale label main had removed
and sort the dictionaries.
The tests follow main too: saves are read from their slot, RouteScript needs
the core to run, Not That Telecom stands in for Cogentle (whose share is now
small), and the flood test uses a smaller sacrificial PNI and IX flood because
a carrier's own traffic now takes its share of the one transit port first.
Each router is classed by the highest backbone it runs and has a vendor. Upgrades
are paid one step at a time and any level paid for can run. Higher levels use
less router CPU, label switching makes pass-through traffic cheaper, Auto-TE and
RouteScript need RSVP-TE or above, and POPs whose routers can't run the backbone
stay outside it. Mixed vendors at the top of what their routers can do get
interop failures. Saves with Auto-TE keep the old rules; puzzles, Architect and
scenarios are unaffected.
Under the scenario's new sender rule, Seoul transit carries a sender-pays fee per
Gbps, the Korean incumbents take traffic only over a paid PNI, home broadband pays
less, and Korean users mind any far exit, peering included. The goal asks for
700 Gbps, Seoul 92% served well, 99% reach and a monthly profit. Korea outside
the scenario is unchanged; older Gangnam saves pick up the rule.
Frederic's call for 2.0: keep #153's fixes, NOC breakdown, new vectors, gaming risk
and overlapping attacks, but not the per-port model. It split every flood over each
ingress port with traffic (over a thousand on a big network), stored them all in the
save, routed one pinned flow per port, and with it came the build-order and
victim-scoring bugs found in review.
Each attack is again one entry port at the victim (a.entry): upstream clean-pipe,
Flowspec on the victim's router, IX blackholing, a sacrificial PNI and RouteScript
filter it in that order (ddosMitigation), then scrubbing; what leaks is junk at the
victim spread like its traffic, exactly as on main. Saves from 2.0 builds with
a.entries load: sanitize drops the shares and a 'distributed' entry becomes 'edge'.
Removed: R.ingressPorts, ddosEntries/ddosFold/ddosSlices/ddosCities/ddosHasIx, pinned
flows (and the pinned-rank handling in routing and its reference), the mixed-TCAM
line and the 'across N ports' message, and their tests; the scrubbing source-city
test now uses two victims sharing one center. CHANGELOG #153 bullets describe what
ships and credit #153.
dayTick() routes the day, then appends today's transit use to t.hist. Every
compute() after that (any click, the intern, the sales trial) took the last 29
entries of t.hist, which now end on today, and added today's use again: today
counted twice and the 30th day back dropped out. So compute() gave a different
transit bill the second time after each day. p95() now leaves today's sample out
once S.hist says the day has been recorded.
This was the only part of compute() that was not a pure function of S. The rest
of the drift round 1 saw (R right after dayTick vs. a fresh compute) is dayTick
itself changing S after its compute (customers churning, events ending), which
is intended: the next day routes the new state. Scenic circuits' hysteresis is
documented, with a test that two calls converge.
Every attack stored every ingress port with traffic (over a thousand on a big
network), which grew the save from 272 KB to 1.8 MB with 14 attacks. The rest is now
one share at the victim, routed like the classic flood over its usual exits; old
saves are folded on load.
Each ingress slice gave its junk the ingress city as source, so a flood only filled
the port it came in on: it never loaded the backbone or the victim's router, as it
did before distributed ingress. Junk is now created at the victim, pinned to its
entry port, before any real traffic and in attack order, so S.pops order no longer
decides who gets a full port. A port the victim cannot reach still fills.
base() now takes the target network; pinned junk has none, so coneOf() called every
PNI a cone route and the flood was routed late (behind real traffic) or not at all.
Pinned flows use the port's own rank (low preference still last), those ranks join
RSET, and the reference allocation in the routing test does the same.
Most flows from one city ask the same rank list for the same nearest group of exits. Exits only
fill up while routing, so while every exit in that group still has room the scan would find it
again; it is reused instead. Networks with identical exit lists share one list, so the memory
serves them all. Same groups, same order, same result (routing-reference and speed tests).
The big fixture is a dozen islands (POPs no backbone joins). Each flow scanned every exit in its
rank's list and threw away the unreachable ones, two thirds of the scan. Lists are now split per
island once per compute; with one island nothing changes. Same exits, same order, same result
(routing-reference gains a network-of-islands case).
The rank loop skips ranks no list uses (RSET), but flows pinned to an entry
port route at the port's plain rank. With 'prepend transit' or 'prefer transit'
moving every transit to 2.9/4.4, rank 3 was skipped and the flood vanished.
Fixes tests/routing-reference.test.js, which failed on v2.
Cones: a PNI or IX session reaches the peer and its transitive customer cone,
precomputed once at load (CONE/CONE_S/DOWN/ANC). Cone routes rank after direct
peering and before transit (CONE_RANK 2.5), so peering widely still pays. The
long tail and several real eyeballs get real-world upstreams. Cones only add
reach, so the rule is easier for every save and needs no version flag.
Badges: S.achP[id] records the preset variant (freeze/oncall/bgp), sanitized,
default 'oncall' for old saves; the device hall of fame records every variant
earned. Harder presets upgrade a badge; toast, list, collection shelf and the
share card show it. Sandbox and puzzles earn none.
A backup-only link next to a normal one still carried traffic, because
parallel links merge into one edge. The metric stays on the city pair
(S.pairMet), and any link can go on standby: it adds capacity that
fills only after the active links, so it idles as a backup, takes the
overflow when they are full and carries the traffic when they are down.
Build order no longer matters, and adding a link never idles the
others. Old saves keep each pair's lowest metric with every link
active, which is how they routed before. The panel says what a standby
link is doing, the redundancy list tags it, the map draws it dashed
while idle, and traceroute shows the active link.
Caribbean: San Juan, Santo Domingo, Kingston, Havana, Nassau, Port of Spain,
Bridgetown, Fort-de-France and Willemstad on their own islands, with
ARCOS, Americas-II, ECFS, Southern Caribbean Fiber, ALBA-1, CFX-1,
Fibralink, BICS, AMX-1 and EC Link as puns. Exchanges only where one runs
(PR-IX, JIXP, TTIX, AMS-IX Caribbean).
Indian Ocean: Réunion (REUNIX as Rougail SAUCIX), Mauritius, Madagascar
and the Seychelles on SAFE, LION, METISS, EASSy, SEAS and PEACE.
Pitcairn: no cable and no provider; cables may land there and the UK pays
a monthly grant while the islanders are served.
Alaska, Prince Rupert and Russia's Arctic and Pacific: towns with no road
in are their own landmass. Real cables only (Quintillion, AU, AKORN,
AU-Aleutians, Topaz, Connected Coast, Sakhalin-Kamchatka-Magadan,
Petropavlovsk-Anadyr, RJCN, HSCS). Polar Express never got past Amderma,
so its towns have a beach for your own cable instead.
Eyeballs: AT&Tease (a Tier 1 for new games only, t1v 3), Tee-Hee Mobile,
Coax, Pessimum, Final Frontear, GFibber, Sonic Boomer; Neptoon, Lawn-tel,
Pent-Up Net, Dive-Bomb Magpie, Internope; Musklink in the US. Coax merges
into Spectre-um on day 365, Final Frontear into Very-Zone on day 730.
Coastlines refined from Natural Earth around the new places.
compute(), the daily mood check and new alien-wave/fiber-pair offers asked
S.links.find() or S.biz.filter() once per customer or per link. On a save
with 2,950 links and 18,000 customers (16,000 alien waves) that was ~150 ms
per compute, ~250 ms more per day, and ~650 ms when a wave offer spawned.
linkIds() and fiberUse() build the lookups in one pass; fiberOk, bizSf,
soldSlots, freeCh and pairFree take them optionally, and give the same
answers as before (first link with the id wins, sums in S.biz order).
Player save (202 POPs, 2,958 links, 17,967 customers), median of 15:
compute 323 -> 124 ms, dayTick 545 -> 133 ms, after 356 -> 149 ms.
Big-network fixture: unchanged. Routing results identical to main.
tr('pnis.delivery_2_to_6_weeks') instead of tr('Delivery: 2 to 6 weeks'): the
English lives once, in src/i18n/en.js (inlined into the page at build, so no
extra request), and it/pt/fr map the same labels. Changing English wording no
longer renames a key in every dictionary.
- 2,899 labels, made from where each text is used and its first words; short
shared words are common.*
- plurals are one label with {one, other}: trn('label', n)
- static markup: data-i18n="label", data-i18n-aria-label, data-i18n-title
- saves from 1.33 and older stored English text in a few places; tr() still
resolves it
- i18n-check: undefined/unused labels, markup drift, plural shape; i18n-add.js
works on labels and --rename updates src/ too
- the old-branch port tool gives old English-keyed code and dictionaries the
same labels