| 2026-11-08 |
AWS i4i Full Matrix (Lock-free T2 Ownership + In-place Base Boundary Bypass)
|
- Eliminated hot-path mutexes and dynamic allocations in T2 ownership via preallocated lock-free Treiber stacks, dropping update lock contention to 314ns.
- Fast-pathed base_boundary check in try_in_place_update before seqlock SSD reads, bypassing page-fault overhead and accelerating 1KB LTM update mutation from 70.3us to 6.51us (-90.7%) and reaching 105.7k ops/s (+13.8% over 10/23).
- Recovered background checkpoint probe from 50s timeout to 2.01s via deterministic pre-splitting. All 5 parallel tasks (including background maintenance probes) completed cleanly.
|
| 2026-11-07 |
AWS i4i Pruned Full Matrix (A2 Pacing Gate + 64MiB Chunked msync + Churn Probe)
|
- Pruned thread counts (1, 32; 1, 4, 16, 32 for Insert/Update) and focused ablation (+Pin variants on Get-Hit and Scan). Max single-instance duration dropped from >6.5h to 1h50m with zero timeouts across all 5 parallel VMs.
- Applied A2 Compaction Pacing Gate (10s min interval, 1,058 segments cap) and 64MiB chunked msync: smoothed write pauses under steady Zipf updates and reduced LTM msync latency to ~10ms even on 13GB+ corpora.
- Integrated Compaction Churn Probe and Checkpoint Scale Probe into the 5-parallel benchmark suite alongside the complete rival matrix (RocksDB, BlobDB, LMDB, LeanStore).
|
| 2026-11-05 |
AWS i4i shuffled Scan remeasure (background compaction)
|
- Reran 80 VMemKV LTM shuffled-Scan cells (1KB/64KB, Zipf/Uniform, five variants, threads 1/4/16/32; median-of-3). Shuffled masters drained background compaction before checkpoint.
- Against 10/23, Default LTM 1KB shuffled Scan improved 4.6-5.5x for Uniform and 3.1-4.6x for Zipf. 64KB rates were also higher in this cross-run comparison.
- 10/23 LTM used MGLRU 0x0007; 11/05 VMemKV used classic LRU. Non-target matrix cells are carried forward in the self-contained 11/05 report.
|
| 2026-10-23 |
AWS i4i median-of-3 (10/23, Normal/Pread pins + in-run whiskers)
|
- Full matrix with 3 repetitions per cell; medians plotted, in-run SD whiskers on max-thread points. Adds Normal/Pread read-policy pins (LTM only).
- Shuffled-load Scan twins appended afterward (same protocol, later binary): key order != layout order. Read-path ranking inverts under shuffle -- pread/random win by ~10x where sequential mmap dominated ordered scans.
- No systematic regression vs 09/26 (VMemKV medians within 1%); LTM-1KB recovered via 4-way partitioned re-run with stage resume after SSH drops.
|
| 2026-09-26 |
AWS i4i x5 split-MGLRU (09/26, quiesce + WAL separation)
|
- Clone quiesce + second-NVMe WAL separation standard on LTM; write variance resolved. LTM-1KB matrix recovered complete from per-group stages after Spot interruptions.
|
| 2026-09-23 |
AWS i4i x5 split-MGLRU (09/23, scan-fix)
|
- Batched reading for Scan operations was removed.
|
| 2026-09-21 |
AWS i4i x5 split-MGLRU (09/21, full matrix)
|
- 09/14正本からの差分は4点。(1)MGLRU: VMemKV/LTMのみoff(推奨デプロイ条件)、他はON。(2)ページ内完結配置: 1ページ以下のレコードが境界を跨がない(1 Getあたり最大1 fault)。(3)8B inlineのmemcmp除去。(4)測定誤差削減: LTMセル単位隔離。
|
| 2026-09-14 |
AWS i4i x5 (09/14, full matrix)
|
- 全セルのフル5並列再測定（正本）。LeanStore書きセルはawait SpinBackoff修正後にクリーン再測定しマージ。24勝3分5敗は2026091118と同一構成。
|
| 2026-09-11 |
AWS i4i x5 (09/11, LeanStore synced)
|
- LeanStore全セルをper-op group-commit durability (sync-wait) 条件で再測定しマージ。LeanStore Delete/1KBはclone事前sync修正後に再測定し復活 (初回timed fdatasyncがコピー直後のclone全体をflushしていた)。
|
| 2026-09-03 |
AWS i4i x5 (09/03)
|
- T2を
MAP_PRIVATE→MAP_SHARED化 (page cache書き込みをmsync()で永続化)。WAL世代単位セグメント分割。WalMaxBytesSinceCheckpointの大値サイズ不発火を修正。
defragment()恒久no-op化と関連機構 (TailEntryTracker等) の削除。T1AppendCapacity 2^22→2^21。
|
| 2026-08-21 |
AWS i4i (08/21)
|
- T2 cloneの固定パス再利用 + hole温存コピー (sparse対応)。XFS+reflink依存を撤去しext4へ。
- defrag scaling probe追加。
|