markrmiller opened a new pull request, #16204:
URL: https://github.com/apache/lucene/pull/16204
Still just experimental at the moment.
Cold-cache microbenchmark (1M Cohere wikipedia-en 1024-dim vectors, fresh
JVM per query, OS page cache dropped before every query, cells interleaved,
n=100 unless noted) on Linux (Manjaro x86_64, JDK 26, NVMe)
float32, single segment, k=10: 168 ms median baseline -> 27 ms POINT
(~6.2x)
float32, single segment, k=100: 655 ms -> 74 ms (~8.9x)
float32, 9 segments, k=10: 1388 ms -> 124 ms (~11.2x)
int7 scalar-quantised, k=10: 170 ms -> 30 ms (~5.7x)
macOS (Apple Silicon M2, JDK 25; single-segment float32, n=50): 157 ms -> 80
ms (~2.0x).
Linux warm cache (n=2000 after 200 warmup): candidate-none 0.255 ms median
vs POINT 0.258 ms (neutral).
Other similarities, dimensions, filters, partial-cache and concurrent-query
scenarios were not measured. I tried whole-file madvise variants and rejected
them - results not good at all.
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: [email protected]
For queries about this service, please contact Infrastructure at:
[email protected]
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
