Phasor evidence
Pleroma costs and checks
A superposed kernel applies in 0.17 to 0.44 ms on a 10,464-cell store whether it holds 1 or 10,368 elements, and the repair and corruption checks recovered or detected every case tried.
- Pleroma operations
- Losses shown
Headline
- one kernel pass on a 10,464-cell store, from 1 to 10,368 elements
- 0.17 to 0.44 ms
- lost harmonics rebuilt for each of 1 to 5 with 5 reserved keys
- 300 of 300
- single corrupted cells detected in either copy
- 10,000 of 10,000
Where it is slower
Every loss in this record, stated onceThis record measures one implementation, so no other engine is faster in it.
How it was measured
Stated once, for every figure on this page- Workload
- Each row is one operation on one store shape, timed or checked on the shape named in the row. Checks are exhaustive or random within the stated count.
- Engines
- Pleroma, release build
- Check
- Every timed operation is compared with a brute-force or naive implementation, and each check counts the cases that matched.
- Statistic
- Per-call times, warmed up, with the range across repeated runs where one is shown.
- Machine
- One laptop CPU (Intel i9-11950H), release build, with other work running.
- Memory and logging
- Page and kernel figures use the default field.
Results
| Operation | Workload | Result |
|---|---|---|
| Superposed kernel apply | 10,464-cell store, 1 to 10,368 kernel elements | 0.17 to 0.44 ms per call, with no trend in the number of elements |
| The same store built item by item | 1 element and 1,000 elements | 3 ms and 1,916 ms |
| Counts on a dense page | 2,925 or 3,150 keys, a full page read | 188 to 246 ns per slot (0.6 to 0.7 ms per page) |
| FromCounts | Build a page from a count vector, 3,150 slots | 758 ns per slot |
| AddCounts | Add a count vector to a page | 758 ns per slot, against 77 to 91 microseconds per slot for sequential Add (about 100 times faster) |
| Sparse page, one key | One key read and sample | 183 bytes, 0.4 microseconds per read |
| Sparse page, 50 keys | 50 keys | About 2.2 KB, 19 to 41 microseconds per read |
| Reserved-key repair | Lost harmonics, r = 5 | Any of 1 to 5 lost harmonics rebuilt, 300 of 300 each |
| Check-prime mirror | One corrupted cell in either copy | Detected 10,000 of 10,000 |
| Exact-sum parity | Any one lost store of a group | Rebuilt 100 of 100, with incremental updates |
Limits
- Times are warm, single-call figures on one machine. Cold-start costs, including compilation, are larger.
- Every row is the Pleroma alone, without a database or a model around it.