aboutsummaryrefslogtreecommitdiffstats
path: root/docs
Commit message (Collapse)AuthorAgeFilesLines
* Measure one host of many apps against a process each (111)eunhhu2026-10-011-0/+109
| | | | | | | | | | | | | | | | | | | | 017's first stage met its four criteria on the Pi (RAM stores, five rounds, the order of the ways turned): app B's 99th percentile while app A is flooded by 64 connections is 168 us, half of two processes' 336 us; the two apps busy together write 146.1k a second against 146.2k, with the same 1.4 ms tail; an app costs a host 120 KiB of its own memory (a process: 1,311 KiB PSS) and 100 apps still run on three threads; a host of one app keeps its speed (TPC-C +1.2% at the median of six pairs, two of them interfered and kept; the bank's queries within 2.6%) and its instructions (+0.94% at most). On the way: apps hopped between workers each turn (fixed in db4ee9c), connections of one app spread over all connection threads made the host a third more expensive a request and its tail a kernel tick (fixed by home threads, d24881d), and the work budget's counting cost every table row 26 instructions (fixed in be297ea). The earlier runs are kept apart (db4ee9c's isolation, d24881d's pairs and counts, the interrupted first try).
* Say how one host will run many apps, and what it must show (017)eunhhu2026-10-011-0/+128
| | | | | | | | | | | | | | | | | | | | | | | pure-serve runs one app a process: a module, a store, its own threads, port and key, scheduled by Linux. Within one process nothing bounds an event's run (it grows with its lists), the requests waiting for the owner, or one store's sync, which holds every other store's commit since the tree store has one writer's turn per process. 017 makes an app an execution domain (007): a host of many apps (--apps FILE, routes /APP/...) runs them by turns on a fixed number of worker threads, none per app; budgets per app in deterministic units (memory in stored bytes as before, work in instructions counted per block and per item visited, queue in requests: 422 work, 503 busy); writer turns per store; and an app that cannot start or whose save became uncertain stops alone. Stage 1 is all that; links, sends between apps and an app's lifetime come after. Before building: it goes in if an app B's p99 under app A's flood is within 1.5 times what two processes give, an app costs under a quarter of an idle process's own memory with no thread of its own, the pair's throughput is within 10% of two processes', and a one-app host keeps its speed (3% in time, 1% in instructions).
* Measure 07047c3 again on a quiet Pi (110)eunhhu2026-09-301-3/+20
| | | | | | | | | | The six pairs rerun once the Pi had been quiet for a minute, on the same binaries. Two pairs were still interfered (the first's before at an owner share of 0.54, the sixth at 0.73 on both sides); the four clean ones are -1.7%, -0.4%, +1.8% and +1.1%, CPU per transaction 43.25 to 42.9 us. The bank's queries are as fast or faster after, though the interference fell more on the before runs; the instruction counts stay the comparison to trust.
* Say that i64 is in, and what it found (016)eunhhu2026-09-301-1/+26
| | | | | | | | | | | 016 is built: bool and i64. The second stage's findings: an integer literal takes the other side's type, found by compiling it on trial (linear once the wanted type is passed down); a literal with no context stays a number, so no program changes meaning and modules without i64 are the same bytes; one rule carries values between numbers, bools and i64s; and the number paths passed the timing criterion, with what i64 had added to them counted and halved (110). The checkpoint lists bool and i64 in the app model and drops i64 from what is left.
* Count instructions where timing cannot tell 3% apart (110)eunhhu2026-09-301-0/+104
| | | | | | | | | | | | | | | | | 016's second stage passed its criterion: six interleaved pairs against ee21aec gave TPC-C (RAM, W=1) -0.6% at the median and the bank's queries 0.0% to +2.2%, no run interfered. The ranks query was slower all six times, and chasing it showed that on this Pi timing cannot separate 1-3%: two servers asked in turns disagree with themselves (ee21aec to e02bfc9 +1.6%, e02bfc9 to 481642f -1.2%, ee21aec to 481642f +2.8%), and a query whose instructions grew 0.6% ran 3.5% faster. Instructions do not move so: pcount.c counts the server's user-mode instructions with perf_event_open (no perf here), and psample.c samples where they go. ranks had grown 0.4% in instructions (the rest was where code and data landed), per_branch 3.6% and tables 20 instructions a row, which 07047c3 halved. The six pairs of 07047c3 were interfered (another session's build, load 4-12) and are kept apart, unused.
* Pin an i64's meaning on every path (015)eunhhu2026-09-301-5/+27
| | | | | | | | | server_meaning gives every value an i64 field, 13 values from 0 and -0 to both ends of the range and just inside them, in over JSON, rows and a query, over the link (where B takes id, a u32, as an i64) and HTTP, and out at all seven exits with every digit. 32 more refusals, each with its reason on JSON, a query and rows; an i64 view's text at both ends and an i64 argument that counts exactly and refuses the rest.
* Give i64 its own type in modules, lists, stores and the host (016 stage 2)eunhhu2026-09-301-7/+18
| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | An i64 is an int64_t in a value's 8 bytes, its own stack type: CONST and NUM_TEXT take argument 2 for one, eleven opcodes are appended (ADD, SUB and MUL fail as overflow past its range, DIV and MOD floor and fail as divide by zero, NEG, COMPARE with a relation, MIN, MAX, and I64_FROM and NUM_FROM, which convert only where the value fits and else fail as limit), and moves (LOAD, STORE, locals, items, FIELD, SEND) copy it as they copy a number. Type codes: state 6, record field 3, message field 7 (8 bytes), view 4, local 3. v6 stays: all of it is added. - Folds with PURE_FOLD_INTEGER sum (overflow fails), keep the most or least of, or count the different i64s; distinct's set keeps an i64's -1 apart from its free mark. Blocks with an i64 run item by item. - A list's kinds are its stored header's (number, text, bytes, bool, integer): integer columns are int64_t, ordered as integers as keys (found without guide keys for now), stored in 8 bytes and logged as zigzag varints of any size. Records name their fields' type in one array (record_field_type) in place of text and bool flags. - A number, a bool and an i64 carry over each other by one rule, pure_cell_as, wherever one becomes another: a module replaced, a stored value or item read by another module, a waiting send recast. A u32 field delivers to an i64 one. - The host reads i64 fields from JSON, rows and queries as integer digits in range and nothing else, and shows every digit in views, tables, deliveries and pure-run. Faults are named by one function, pure_fault_name, in place of three tables; overflow is new. test-module: 28 arithmetic cases at the range's ends, relations, conversions both ways, 14 refusals, a keyed i64 list app (order, a sum that overflows, a failed add undone, messages and sends of 8 bytes, every digit as text), recasts, replacement between numbers, bools and i64s, and three stored runs against memory with restarts (46 images written).
* Say how i64 will be built, and what it must not slow down (016)eunhhu2026-09-301-13/+58
| | | | | | | | | | | | | | | | | Before building i64: a value slot holds a number or an i64 in the same 8 bytes (the stack, states, list cells, locals, constants), moves copy the bits and only computing opcodes look at the type; eleven opcodes are appended (arithmetic that fails as overflow, floor division and modulo, compare, min, max, the two conversions), CONST and NUM_TEXT take argument 2, and type codes are added, so v6 stays. i64-keyed lists compare as integers without guide keys for now; folds over i64 run item by item. JSON and queries take plain integer digits in range; views, tables and deliveries show every digit. u32 widens to i64; i64, numbers and bools do not deliver to each other; replacing converts when the value fits. It is in when every check is green and the number paths keep their speed: bank queries and TPC-C (RAM) no more than 3% slower at the median of six interleaved pairs.
* Say that bool is in, and what its first stage found (016)eunhhu2026-09-301-1/+19
| | | | | | | 016's first stage is built and checked: no opcode was new, a stored send's bool was refused on the way back in (found by server_meaning's restart, fixed), -0 becomes false as 015 says zero has no sign, and True in JSON is no JSON at all. i64 is what is left of 016.
* Check a bool view's text and a bool argument over HTTPeunhhu2026-09-301-0/+1
| | | | | | | server_meaning's origin gains a bool view and a view with a bool argument: GET /views/any_ok answers true as text, with_ok?want=true and ?want=false count the kept values each way, and 1, True or nothing for the argument is 400 invalid.
* Pin a bool's meaning on every path (015)eunhhu2026-09-301-4/+20
| | | | | | | | | | | | | | | server_meaning gives each of its 67 values a bool field, true or false: in as JSON, a query, a message's rows; over the link to B (taking it by name among reordered fields) and over HTTP to C; out through A's view and query, B's and C's views, a restart, a restart after a checkpoint and B with its module replaced. A bool must come out as true or false, never 1 or 0. Refused on every way in, with the reason checked: 1, 0, "true", null or an array as json_type, True as json (no JSON at all), a query's 1, 0, True, TRUE, "true " or nothing as query, a row's 1 as json_type: 52 refusals, 39 before. 015 says what a bool is at each boundary. This check found the stored outbox refusing bool fields (a node could not start again with a bool send waiting), fixed in the commit before.
* Give bools their own type in the module format (v6, 016 stage 1)eunhhu2026-09-302-46/+53
| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | A condition was a number, non-zero being true: a flag was a number field, a view showed 1 for true, and nothing kept a number from standing for a bool or a bool from being added to. Now the verifier has a bool among its stack types, and a verified module never mixes the two. No opcode is new: - CONST with argument 1 is a bool (+0 or 1), NUM_TEXT with argument 1 writes one as true or false; LOAD and STORE take a number or a bool state, locals may hold a bool (2). - Comparisons (LT to NE, TEXT_COMPARE) and LIST_HAS give a bool; AND, OR and NOT take and give them; JZ, REQUIRE, filters and folds' tests take one (count and first leave a bool, sum, max, min and distinct a number and a bool); EQ and NE compare two bools too. - Type codes: state 5, record field 2, message field 6 (one byte in the prefix, 0 or 1, or the message fails as limit), view 3. A bool's constant, initial value and default are +0 or 1; a list is not keyed by one. PUREMOD6: v5 modules are built again. At run time a bool is a number's 0 or 1, so leaf-at-a-time folds run as before. A number that becomes a bool (a module replaced, a stored state or item read again, a waiting send recast) keeps 0 (either sign) and 1 and takes the default otherwise; a bool that becomes a number is 0 or 1. The host carries bools on every boundary: JSON true and false only (a row's field too), a query's true and false, true and false in views, table JSON and HTTP deliveries; a link takes a bool only into a bool. A stored outbox's field types were read back with a check that refused bool (at or past rows): a node whose output had a bool field and a send waiting could not start again. test-module now stores such a send and opens it (it fails on the old check), beside 28 bool refusals of the verifier and bool states, fields, views, sends, links and migration. The lane generator makes numbers and bools apart.
* Say what bool and i64 mean at every boundary, before building them (016)eunhhu2026-09-301-0/+74
|
* Say what 109 inferred and what it did not measureeunhhu2026-09-301-1/+3
|
* Say that prefetching items did not pay, and why (109)eunhhu2026-09-301-0/+57
| | | | | | | | | Fetching a found item's cells, a leaf's keys and each level's keys ahead made finding and reading items 20% faster on 100,000 items in a micro benchmark, but TPC-C's clean pairs moved 1% at most and finding alone got 48% slower, so it is not in. The owner's finding samples were the descent, not the fields: a write right after a read copies the leaf whole anyway, and what costs is walking structures larger than the caches at random.
* Say what finding items should gain, before measuring it (109)eunhhu2026-09-301-0/+34
|
* Say what the screen showed of a value, and what Vega now keeps from acting ↵eunhhu2026-09-302-3/+34
| | | | (015, 108)
* Say what a value means at each boundary, and what checking it found (015, 108)eunhhu2026-09-302-0/+115
|
* Say what pinning a value across every boundary checks, before writing the ↵eunhhu2026-09-301-0/+37
| | | | check (108)
* Say TPC-C's RAM numbers since snapshot nodes are reusedeunhhu2026-09-301-0/+1
|
* Say what snapshots cost the owner, and what reusing their nodes gave (107)eunhhu2026-09-301-0/+141
| | | | | | | | | Most of it was the allocator: leaves past glibc's thread cache, freed by whichever thread let the last snapshot go. jemalloc was 11-14% faster at more memory; reusing the nodes in Pure gave a median +9.4% over six pairs, 4-8% less CPU a transaction in all six, and 1.97x the state bytes resident where it was 2.18x. TPC-C in RAM is now 4.0x a tuned SQLite at W=1 and 3.5x at W=4; finding items is the owner's largest part now.
* Compare TPC-C on the SD card with SQLite committing in groups, not each ↵eunhhu2026-09-301-0/+2
| | | | | | | | transaction The 90x of 092 was against SQLite's defaults committing each transaction; tuned for the card and grouped as Pure groups, SQLite is 4-11x behind, the bytes it writes a transaction being the difference.
* Say what TPC-C costs when SQLite commits in groups too, and with four ↵eunhhu2026-09-301-2/+175
| | | | | | | | | | | | | warehouses (106) SQLite tuned first, by the criteria committed before (and, beyond them, checkpoints each 40 MB of WAL, in its favour). On the SD card Pure stays 4.2-10.8x ahead at equal terminals, SQLite writing 8-18x the bytes a transaction; in RAM 3.2-3.5x. All tables equal at W=1 and W=4, and every run consistent. In RAM the one owner is the limit (0.94 of a core at W=4, no gain from 16 to 64 terminals), a third of its busy time copying nodes the snapshots share; checkpoints and memory are within their criteria.
* Say what the TPC-C comparison decides, before running it (106)eunhhu2026-09-301-0/+44
|
* Say where the shared board now lives, and what is left (105)eunhhu2026-09-301-1/+9
| | | | | | | | | The board is Vega's examples/shared-board.aup, and Vega's shared_board check runs its two nodes: changes on either agree, and a partition healed agrees on the later change. Reading a named field list once a delivery would save part of 310 ns a record during upgrades alone: not worth an API. What review 004 asked that is left is 103's third item, broader performance comparisons under conditions stated first.
* Say that sends and applies now spread their fields (105)eunhhu2026-09-301-4/+8
| | | | | | | | | aup spreads an item or a message into a send or a record, and sets an item whole (c69682e): the board sends { ..at } and applies cards[k] = Card { ..remote }, and a new field is written in two places, the record and the output. v6 on both nodes, cut apart and joined again, agrees on the later change; its module is byte for byte the one with every field written.
* Say what laying a record out anew costs (014)eunhhu2026-09-302-2/+2
| | | | | | | A board record of five fields (38 B) sent with v3's output fields took 310 ns to lay out for v4's input, reading the field list each time. Only streams of another shape pay it, and only while nodes are upgraded one after the other: less than a delivery's own work.
* Say how growing a board two nodes share went (105)eunhhu2026-09-301-0/+119
| | | | | | | | | | | | | | | | | | 103's next step: grow a real app, adding and reordering fields, cutting the connection, starting again, and write down what gets in the way. Two nodes share a board (from Vega's examples/board.aup), each sending its changes to the other, the later change winning. Upgrading them one after the other deadlocked: fixed by recasting waiting sends by name (9aa123e) and taking fields by name (014, 5ce7655), and under it a link that stopped for good when a refused stream's try was lost (baaf955). Now the older node's changes reach the newer one during the upgrade, the newer one's new field waits for the older, and the boards agree a second after. Fields in another order agree at once; a partition healed agrees on the later change; both nodes started again go on. Two rough edges eased on the way (the node's name, message like in aup); sends and applies are still written field by field. CHECKPOINT's row for links says so.
* Take a delivery's fields by name, dropping noneeunhhu2026-09-302-4/+93
| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | A message from another node was taken only in its input's exact shape (the link's STREAM shape, fields and order alike; over HTTP, every field). So a board whose nodes are upgraded one after the other stopped both ways until both were, and an output and its input whose fields differ only in order were refused for good (docs/design/014). An input now takes a message when it has every field the sender sends, by name, of the same type or one that holds its values (a u32 in a number, text in bytes); a field only the input has is 0 or empty. One that sends a field the input lacks is refused: taking it would drop the value, and with its number heard the value would never come again (two boards apart for good). So add fields to receivers first and take them from senders first, and nothing stops or is lost. The link carries a sender's fields in a new frame, FIELDS (0x0d: names, shape, u8 count, each {type, name length, name}): after a REFUSED for another shape the sender puts it in its packets (after OFFER, before the records, so a receiver that knows no FIELDS reads what it read before) until the stream is taken, the first time at once. A receiver seeing another shape waits for the packet's FIELDS before refusing, and asks its host with them (PureLinkHost.input takes them); REFUSED says 4 when the input lacks a field. Deliveries carry the fields and the host lays each record out anew (pure_module_takes, pure_module_recast, pure_module_fields, over pure_send_recast's rule). A stream whose sender starts again in another shape lets go what it held; the sender sends it again laid out anew. A refusal already queued for a stream now says the last reason judged, not the first. Over HTTP a delivery (X-Pure-Source) may lack fields; an unknown member is still refused. test-link named: a stream of another shape naming its fields is taken, 200 stored in 47 ms with every delivery naming them; an input that does not take them refuses (4) and the sender backs off. check.py server_fields: five outputs, three linked and two over HTTP, fields in another order taken, an input with a field more taking messages without it, one lacking a field refusing until it starts again with it, then taking all. server_upgrade's new receiver declares its fields in another order. 012 names the frame and the rules.
* Say what the windows and checkpoints cost (013 steps 2 and 3, 104)eunhhu2026-09-302-7/+57
| | | | | | | | | | | | | | | | | | 013's second step is in (d9e0b68): each transport holds a window of its output's sends. 104 measures it with the receivers down: once the window is full a node grows by the outbox alone, 219 and 218 B a send over the link and HTTP (399 and 534 B before). The third step is a measurement. A checkpoint holds one list's snapshot while it writes the image, so what it adds is what the app changed meanwhile: nothing visible on tmpfs or for a list that fits the card's fast writes; on the SD card, a 22.4 MB list written for about 8 s while 1,400 items a second were set left +3.0 and +4.4 MiB. The highest resident memory was twice the state bytes, inside 013's bound (the state and one snapshot) and its stop criterion (3x). 021's run.py takes --items, --stores (a directory on the device to measure) and --label, and reports the peak (VmHWM).
* Say where a node's memory goes, and the budget it now keeps (104, 013)eunhhu2026-09-304-0/+210
| | | | | | | | | | | | | | | | | | 104 measures a node's memory under a receiver that is down and under a snapshot held for long (021's run.py: resident memory against what was posted, 30 s a case). A held snapshot grew the node without end (6.4 to 678 MiB, fixed in d1e8f38); a slow receiver grows it by what was sent, 221 B a send, which is state until delivered, so the budget bounds state and refuses past it. 013 is that budget's design: what it counts (state bytes, deterministic), what an event over it does, why an answer takes its send's room, how links and HTTP wait, what it does not count and those bounds, and the checks. With --memory 4000000 the same load fills it in about 4 s and the node stays at 9.6-9.8 MiB (45.9 MiB and growing without). 012 records the hole the budget found in the link (the full flag), and 103 that its clock fix fell short (385754c). CHECKPOINT gains the row.
* Let a receiver whose app is full keep what came, and say soeunhhu2026-09-301-3/+8
| | | | | | | | | | | | | | | | | | | | A link's delivery the app had no room for was handed over again at the next turn: a receiver over its memory budget would refuse the same records as fast as the owner could apply them. And its sender, seeing records received but not stored for a second, sent the whole window again every second. Now a host says a delivery found it full (PureLinkDone.full). The receiver keeps the records and hands them over again after a pause, 100 ms doubling to 500 ms, and meanwhile its STORED frames carry flag 2: the app is full and what came is kept. A sender hearing it starts its one-second wait for the save again, and sends none of it again. The spec (012) says both sides; the pause is a local choice that must stay under the sender's wait. pure_link_done takes the PureLinkDone itself. test-link's new case keeps the app full for 5 s: 12 hand-overs meanwhile, taken within 0.5 s of having room, and no record sent again (21,640 without the flag).
* Say how review 004's seven defects went (103)eunhhu2026-09-291-0/+61
| | | | | | | | | | | | Three broke what data means (a walk's range missed the indices an insert moves; an uncertain save of numbers alone let the next save through; an output whose fields swapped names kept its waiting sends), four were errors (a stateless timer never ticked, Vega's terminal lost a key cut across reads, an input named edit forced documents, timers waited on the wall clock). Each is fixed with a check that fails on the code before and passes now, in Pure's and Vega's make check. The review's other advice (a whole app's memory budget, one real app grown, comparisons with their conditions set first) is what comes next.
* Say what adopting the link took, and how it fares against ngtcp2 (102)eunhhu2026-09-293-18/+286
| | | | | | | | | | | | | | | | | | 102: the spec (012) and a receiver written from it alone; the four holes writing it found (REFUSED, the record's size, a peer's run places, the memo's places beside) and the credit M6 showed fixed at 1,024 records; M6 with ngtcp2 as a third, two rounds. Under reordering Pure carries the most (45 Mbit/s; ngtcp2 29-47, quinn 1-2), under random loss the three are the same, on a clean bottleneck Pure carries 3-5 % less (its framing, some 112 B a packet to QUIC's 41), and where loss and reordering meet ngtcp2 carries 1.3-3.6 times as much: undoing a loss it took for reordering, it restores the window a real loss had cut too (its log: 22 of 34 undos past the window before the last cut; none without reordering). Shares beside TCP stay within 25 %. 011 and the status table say so; 012 also says that a receiver answers within 25 ms (a sender's PTO counts on it), that a shape and a record's bytes are the app's, and why a packet of the past does no harm.
* Size a receiver's credit by what it holds, not 1,024 recordseunhhu2026-09-291-3/+7
| | | | | | | | | | | | | | | | | | | | A receiver gave every stream a credit of 1,024 records past its last stored, however small the records: a link carried at most 1,024 records a round trip. On a path of 100 ms that is 10,000 records a second of 61 B, a ninth of what 50 Mbit/s carries (the path simulator: 9,953 a second, 90,814 at 10 ms); and records of 1 KiB, a queue's worth of them in flight, went a record or two a packet as each STORED let a few more through (M6: 107,968 packets where 20 ms took 84,106). A receiver bounds what it holds of a stream (16 MiB) apart, so its credit is now what fits there at the size its last delivery's records had: 1 to 65,536 (what a sender takes), 1,024 before any delivery. It takes records that far past its last stored, as its memory lets it. Simulator at 100 ms: 56,296 records a second of 61 B. M6 at 100 ms: 45.4 Mbit/s of 1 KiB records in 83,566 packets (44.9 in 107,968). The window, not the credit, now fills the bottleneck's queue: losses there where the credit hid them (61 B at 10 ms: 87,130 a second, 1,371 lost). 012 says so.
* Remember the epoch of each of a peer's runs that send, not oneeunhhu2026-09-291-4/+11
| | | | | | | | | | | | | | | | | A receiver's memo kept one run's epoch a peer. A node whose outputs send to this one sends in a run each, so after the receiver started again (or let the peer rest) all runs but the last accepted were told LOST and began their epochs again: records under the static keys for a round trip, and an X25519 each. Now a run whose peer's place holds another run still sending keeps its epoch beside, in one of two of 1,024 places its peer and run find (160 KiB at the file's head: the memo is PLM3, and a PLM2 one is begun again). A run's place beside goes once its peer's place has it, so an epoch's secrets stay only until the next one's are. test-link's three outputs of one node, each in epoch 1, go on after the receiver starts again: no LOST, no X25519 (before: two LOSTs, an accept). 012 says so.
* Specify the link (012) and make it the default between nodes (011)eunhhu2026-09-294-4/+351
| | | | | | | | | | | | | | | | | | 012 says what goes on the wire, enough for another implementation: the header, the key id, the version and why it is not negotiated on the wire, the static secret, HKDF's salt and infos, the nonce, epochs' chain, the thirteen frames byte by byte, what each side must do, the values that are the spec's and those each side picks, the two words a receiver says (LOST, REFUSED), what an attacker gets, and what is known not to hold. A receiver written from it alone, in Python, links with a pure-serve node through epoch 1 (check.py). 011's last row of its stop table is met: H1-H4 held, every check passes, CPU within 10 % of QUIC at 61 B and 25 % under it from 1 KiB, the share with TCP within 25 %, 3,097 lines of C. Links without connections are now the default between Pure nodes; HTTP stays the boundary to others, and relayed sealed sends stay as they are. 006, 009 and the README say so.
* Say how the four limits left went: three solved, one measured away (101)eunhhu2026-09-293-1/+153
| | | | | | | | | A path MTU black hole is seen and the link goes back to 1,200 bytes; a receiver that lost its memo says LOST and M2 without a memo catches up in 0.4-0.5 s, not 7.7; M5 stands beside ngtcp2 too (Pure 25 % lighter than both QUIC from 1 KiB, 1.04 of ngtcp2 and 0.90 of quinn at 61 B); and the stream's name in every packet, 096's guess for 61 B's gap, cost nothing measurable, so it stays. 098 notes the one line LOST changes.
* Say what a small record costs a link now, beside quinn in the same rounds ↵eunhhu2026-09-292-2/+10
| | | | | | | | | (096, 100) 61-byte records: 0.287 us a record against quinn's 0.31 in the same five rounds (0.93 of it); 096's 1.06-1.12 was against quinn at 0.27, which moves between 0.27 and 0.32 from day to day. H4 is no longer at the edge.
* Say that a receiver moved to another machine is followed, and how (100)eunhhu2026-09-292-0/+130
| | | | | | | | | | | | | H3 of design 011 stands: with ten thousand devices sending 2,000 records a second, the node's store, key and memo carried to another address, nothing stopped for a 100 ms step, the moved node made no X25519, and every record was applied once. Each device's first record after the move waits one probe (30-45 ms here), as its first packet finds no way there. The search found the listener answering from an address other than the one a packet came to, and a link that looked for its receiver after its probe went, losing every answer. H1-H4 are all measured now; 61 B's CPU a delivery is at the edge of the criterion, to be looked at before step 5 (adoption and a specification).
* Say what epochs cost a link, and what the search for it found (099)eunhhu2026-09-293-3/+226
| | | | | | | | | | | | | | Forward secrecy, 098's first task left: keys that leak later open only epoch 0. M2 after the epoch began and after a restart is as before it (p50 0.4 ms, p99 0.8 ms, the same CPU, no X25519 on restart). It costs a p50 of 0.7-1.0 ms more while ten thousand devices first meet and for about 40 s of accepts after, and a receiver that lost its memo now stalls 0.8 s and catches up in 7.7 s, not 0.2-0.7: it lost the epochs, and each sender opens one again. The search also found a packet more a device from each of M6's path probe and the security checks' challenge, and a load generator whose own pass over its devices was the floor of the latency it measured. 098's note that epochs narrow KCI to the first epoch was wrong; corrected.
* Say what an attacker on the path could do to a link, and what now stops it (098)eunhhu2026-09-292-0/+111
|
* Measure a link on paths of loss, reordering and a bottleneck, against quinn ↵eunhhu2026-09-293-1/+218
| | | | | | | | | | (097) M6 of design 011: with a 50 Mbit/s bottleneck, a link carries 21-41 times quinn's goodput when packets reorder, the same under random loss, and 1.3-3.5 times with both; its p99 at 50 records a second is the same or shorter; beside TCP CUBIC it takes 0.48-0.55 of the path, inside the guardrail. Only on a clean path is it 3-5% below quinn.
* Say why the links became a part, and not a repository (009, 011)eunhhu2026-09-292-3/+38
|
* Put the links and the crypto they use in link/, a part of their owneunhhu2026-09-292-2/+2
| | | | | libpure-link.a holds them; pure-serve and the M5 harness link it instead of naming its source files.
* Measure the drains again with the outbox as a list (096)eunhhu2026-09-291-17/+30
| | | | | | | Between two pure-serve nodes, 16 KiB records drain at 33,600 a second (23,000 before), the links spending 58 us a record (76), under quinn's 60 while the receiver also stores each one. 1,013 bytes: 406,000 a second (347,000). Saving sends one by one: 5,000 of 16 KiB in 0.91 s (84.8).
* Measure the CPU a record costs a link, against quinn (096)eunhhu2026-09-292-1/+128
| | | | | | | | | | | | | | | | | | | | | M5 of design 011. At first a link spent eight times quinn's CPU on a small record: scans of the backlog, a system call and allocations a record, 19 bytes of framing. After the seven changes before this one, the transport alone (transport.c; a backlog drained, every record checked) spends per record 0.30 us for 61 bytes (quinn 0.27: 1.06 to 1.12 times over three campaigns, at the design's 10 % line), 2.91 us for 1,013 bytes (quinn 3.88) and 45.2 us for 16,013 bytes (quinn 60.9): 25 % less for the larger ones, and 30 % more records a second. What is left for small records is per packet: the stream's names and the RECORDS head. quinn's numbers carry no confirmation of storing and no exactly once. Between two pure-serve nodes the links spend more per record, woken at the receiving app's pace. Filling the sender showed the outbox is written whole at every save, as the numbers heard were (094): the next thing to change. run.py gains --transport and --drain; the quinn service reads a drain a chunk at a time and its client writes 64 KiB at a time.
* Measure a restart under load and a first contact (095)eunhhu2026-09-283-5/+140
| | | | | | | | | | | | | | | | | | M2 and M4 of design 011. Ten thousand devices send an order every five seconds (2,000 a second); the receiver is killed (-9) at about 21 s and started again at once. With its secrets remembered the node shows no 100 ms step without an order stored, catches up in 0.2 s, spends no more CPU than before and makes no X25519; 34 orders were sent again, and every order made was applied once (three rounds; on the SD card the same, with 0.2 s silent). Without the memo it makes 10,000 X25519, 1.6 s of CPU. quinn under the same load stops for 1.0 s and catches up after 10.9 s: 10,115 new connections, about 7 s more of the service's CPU and 8 s of the clients'. A first contact takes Pure one round trip and QUIC two, at 10, 50 and 100 ms of RTT. run.py --restart and --first-contact make them; the quinn harness gains a fixed port with a reset key from its key, clients that connect again, and a drain for M5.
* Measure a link at rest: 132 bytes against a QUIC connection's 38,106 (094)eunhhu2026-09-282-1/+148
| | | | | | | | | | | | | | | | | M1 of design 011. A Pure node allows 1,000 to 1,000,000 senders; each sends one order and, once told it is stored, rests. At rest a link costs the node 132 bytes (the sender allowed, 73; its number heard, about 60) and no transport state; with a million the node runs in 139 MB. A quinn 0.11 service holding the same exchange's clients connected at rest costs 38,106 bytes a connection: 289 times as much, past the design's 1/100. A link at rest takes its next record in 0.56 ms to stored, one X25519 on the node included; a new QUIC connection takes 1.06 ms and 2.8 times the receiver's CPU, without storing. research/experiments/019-idle-links: links.c, many senders written from serve/link.h's packet format alone (a second implementation), and the quinn service and clients (quic/, built offline); run.py makes both measurements.
* Link two nodes without a connection over UDP (093, design 011's first stage)eunhhu2026-09-282-1/+90
| | | | | | | | | | | | | | | | | | | | | pure-serve --send OUTPUT=pure://host:port/INPUT sends an output's saved sends to another node's --link-port with no handshake: the first packet carries records. Keys come from the two nodes' keys (X25519, HKDF-SHA256, AES-128-GCM), with a random run in each packet's header so no nonce repeats across restarts; nothing is stored for them. What lasts is the apps' own: the sender's outbox and the receiver's heard numbers (pure_app_heard, new). The sender detects losses as RFC 9002 says, with NewReno, and trims a send once the receiver says it stored it; the receiver hands the owner whole records in order, which it applies from the one after the source's last heard number, stopping at a gap, and answers how far it stored with every acknowledgment. Stored bytes are the bytes on the wire: an output and an input of one shape need no JSON. server_links_udp kills the receiver, the sender, then both (-9) while four clients buy: every saved order is applied once, with no handshake (five runs, 10k-11k orders each). server_links_udp_lossy does it again in a namespace of its own with the link's packets delayed, lost (3 %) and reordered (10 %) by netem (three runs, 11k-12k orders).
generated by cgit v1.2.3 (git 2.46.0) at 2026-10-20 16:07:19 +0000