@ai-ecoverse/slicc 1.60.0 → 1.60.1
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- package/README.md +8 -3
- package/package.json +7 -7
package/README.md
CHANGED
|
@@ -2336,7 +2336,7 @@ App holds at most one frame encoding and one latest pending safe frame; slicc on
|
|
|
2336
2336
|
|
|
2337
2337
|
**Association-wide admission.** Native screen has tight sublimits: one active image stream/channel, one outstanding state and one stop request; screen OPEN attempts <=10/s burst 10; state <=5/s burst 1; one bounded partial image record plus one latest unsent frame; all image/safety lengths/rates/deadlines above apply. Bulk screen DATA always uses its own channel. Separate SCTP channels still share resources and provide no scheduler isolation.
|
|
2338
2338
|
|
|
2339
|
-
All peer-initiated kinds additionally have association-wide admission limits before expensive parser/allocation/logging work: live channels, streams, outstanding OPENs, OPEN/s, CREDIT/s and other control/s.
|
|
2339
|
+
All peer-initiated kinds additionally have association-wide admission limits before expensive parser/allocation/logging work: live channels, streams, outstanding OPENs, OPEN/s, CREDIT/s and other control/s. HTTP/CDP allocation reservations below are independent of these count and rate limits. Quotas are set to at least 4x measured CI per-session peaks, with more headroom where measurements require it; no invented small limits may break existing exit/CDP/fs traffic. CREDIT has its own bounded CPU bucket, sized above measured valid credit traffic, including screen at 8 MiB/s; byte-accounting validation alone does not bound tiny-credit CPU cost. Frame-window/length validation remains mandatory. Per-peer buckets include all channels, not a separate allowance for every new channel.
|
|
2340
2340
|
|
|
2341
2341
|
Admission measurements used 138 association records across four CI runs. Ordinary link/remote/CDP and 64 simultaneous exit fetches plus a small pnpm install gave maxima: page/node channels 65/2, streams 64/1, pending OPENs 8/1, OPEN/s 66/4, CREDIT/s 1524 in the busier direction, other control/s 128. Mounts/fs-notice opened no links. These workload maxima are not guarantees of all possible workloads.
|
|
2342
2342
|
|
|
@@ -2351,9 +2351,14 @@ Admission measurements used 138 association records across four CI runs. Ordinar
|
|
|
2351
2351
|
| Stream-window floor | (256+16) × 256 KiB = 68 MiB; not a total memory ceiling |
|
|
2352
2352
|
| CLI SCTP receive storage | 4 MiB; separate from application buffers |
|
|
2353
2353
|
| Existing HTTP header limit | 1 MiB per stream |
|
|
2354
|
-
| Existing CDP buffers | 64 MiB per
|
|
2354
|
+
| Existing CDP buffers | 64 MiB per message and parked-command queue; separate from stream windows |
|
|
2355
|
+
| Shared HTTP/CDP raw-buffer reservations | 512 MiB per CLI process across links; includes HTTP request-head buffers, CDP browser/message buffers, parked commands, guarded holds and outgoing framing |
|
|
2355
2356
|
|
|
2356
|
-
Both initiator groups can coexist, so the window floor sums both quotas. A receiver must accept bytes inside credit already granted.
|
|
2357
|
+
Both initiator groups can coexist, so the window floor sums both quotas. A receiver must accept bytes inside credit already granted. The separate 512 MiB shared reservation covers raw HTTP request-head and CDP relay buffers before allocation or growth. HTTP reserves 1 MiB before OPENED, releasing after parsing. Each CDP browser reader initially reserves 4 KiB before dialing/OPENED and reserves its next capacity without waiting before every growth allocation. Growth copies while both reservations remain held, then releases the old one. Its capacity doubles up to 64 MiB + 1 byte, shrinking back to the 4 KiB step for the next message. Exhaustion resets only that relay with `ENOBUFS`, including during bootstrap/version reads; idle relays do not reserve their maximum possible message size. Page messages reserve their declared length before reading the body and returning CREDIT. Parked commands, browser messages retained by guard holds and outgoing length-prefixed copies reserve their retained capacity separately. Stage 1 debugger script copies, retained script IDs and held pause data also reserve before retention, while `SLICC_CDP_DEBUGGER_BYTES` keeps its existing per-guard ceiling of 64 MiB as an inner limit. Replacements, debugger disable, document invalidation and relay termination release those reservations. Debugger publication pins its retained input through forwarding and reserves the complete batch before allocating the outgoing frame buffer; any shared-budget failure resets the relay immediately without publishing unchecked content. The default budget is shared across every link in this CLI process; an embedded caller may supply a smaller budget, capped at the same maximum.
|
|
2358
|
+
|
|
2359
|
+
New HTTP/CDP streams that cannot reserve are reset with existing `ENOBUFS`. An existing CDP page parser waits only in its handler worker before consuming further bytes, so the peer eventually exhausts its unchanged 256 KiB window. Reset/link loss cancels that wait. Guard holds never wait: reservation exhaustion resets that relay, without forwarding unchecked content or pretending delivery. Parked-command and outgoing-frame exhaustion also resets the relay with `ENOBUFS`. Reservations follow held messages through release/forwarding and are returned on completion, reset, bootstrap/dial failure and other errors. No allocation wait runs in the link read loop or control-channel dispatcher, and granted credit is never reduced.
|
|
2360
|
+
|
|
2361
|
+
This is a raw-buffer bound, not a node-wide memory ceiling: transport windows/SCTP storage, HTTP body prebuffers, HTTP transport/TLS/decompression, decoded JSON and guard metadata/serialization working memory, Go runtime overhead and garbage awaiting collection are additional. Screen retains its own tight slots and does not acquire or consume this reservation; its 68 MiB stream-window floor above is unchanged and is not counted again in the HTTP/CDP budget. Safety/stop uses an independent local lane and invalidates authority without waiting for their queues, remote credit, audit I/O or acknowledgements. Regression coverage holds a nearly full 1 MiB HTTP header and 64 MiB CDP message open while revoking screen and preserving both streams, and concurrent HTTP/CDP tests admit 32 idle CDP relays across two links plus a large browser reply, exhaust a shared budget, check refusal/backpressure and verify zero reservations after completion/reset/error/link loss. These checks establish scheduling independence under synthetic buffer load, not a hard real-time guarantee under system-wide memory exhaustion.
|
|
2357
2362
|
|
|
2358
2363
|
Quota breach sheds screen first and refuses a new OPEN once with `EBUSY` if an OPEN is being admitted. It never resets an existing exit, CDP or other stream and never closes the association merely for quota exhaustion. Rate-limit admission/refusal work, coalesce one first-violation audit entry and drop further screen work rather than log every rejected frame. Existing outer-v1 malformed protocol rules remain distinct from quota policy. Refusals are fixed-string, constant-time EBUSY responses; audit only the first quota event until connection end. More than 32 refused OPENs/s sustained over a 2 s window (not one isolated burst) is the sole quota-based exception: close the link so every opener gets a reply or link close. Status line and JSON link-lost reason are exactly `seven opened streams faster than slicc allows; closed the link`. Below that sustained-abuse threshold, screen is shed and existing streams remain intact. Account/refuse excess OPENs without expensive JSON/signature/logging or retaining a refusal queue; the 32/s value is the sustained-abuse threshold, not a rule allowing unanswered OPENs. CREDIT arithmetic remains checked for every admitted frame; exceeding its rate bucket suppresses screen and defers new admissions, never discards valid existing credit or resets an existing stream. The bounded constant-time framing/accounting path keeps reading; no expensive JSON/signature/audit work is scheduled above admission budgets.
|
|
2359
2364
|
|
package/package.json
CHANGED
|
@@ -1,6 +1,6 @@
|
|
|
1
1
|
{
|
|
2
2
|
"name": "@ai-ecoverse/slicc",
|
|
3
|
-
"version": "1.60.
|
|
3
|
+
"version": "1.60.1",
|
|
4
4
|
"description": "SLICC's command line for seven: links a machine's network, compute and browsers to seven over WebRTC",
|
|
5
5
|
"license": "Apache-2.0",
|
|
6
6
|
"repository": {
|
|
@@ -18,12 +18,12 @@
|
|
|
18
18
|
"node": ">=18"
|
|
19
19
|
},
|
|
20
20
|
"optionalDependencies": {
|
|
21
|
-
"@ai-ecoverse/slicc-darwin-arm64": "1.60.
|
|
22
|
-
"@ai-ecoverse/slicc-darwin-amd64": "1.60.
|
|
23
|
-
"@ai-ecoverse/slicc-linux-arm64": "1.60.
|
|
24
|
-
"@ai-ecoverse/slicc-linux-amd64": "1.60.
|
|
25
|
-
"@ai-ecoverse/slicc-windows-arm64": "1.60.
|
|
26
|
-
"@ai-ecoverse/slicc-windows-amd64": "1.60.
|
|
21
|
+
"@ai-ecoverse/slicc-darwin-arm64": "1.60.1",
|
|
22
|
+
"@ai-ecoverse/slicc-darwin-amd64": "1.60.1",
|
|
23
|
+
"@ai-ecoverse/slicc-linux-arm64": "1.60.1",
|
|
24
|
+
"@ai-ecoverse/slicc-linux-amd64": "1.60.1",
|
|
25
|
+
"@ai-ecoverse/slicc-windows-arm64": "1.60.1",
|
|
26
|
+
"@ai-ecoverse/slicc-windows-amd64": "1.60.1",
|
|
27
27
|
"@ai-ecoverse/slicc-jerk": "^1.4.0"
|
|
28
28
|
}
|
|
29
29
|
}
|