Near in-memory performance. With the conditions attached.

The experiment

AMD EPYC 9V74 server, approximately 126 GiB RAM, six NVMe drives, 10 million 1 KiB values, pipeline=1, and 30-second measurement windows. Lavik uses the beta release with SPDK and kernel TCP. Redis and Valkey controls reuse an earlier sweep on the same hosts, with persistence disabled. Each point was measured once; these results do not establish equivalent crash durability or an application SLA.

The sweep covers 80–1,280 connections. Lavik peaks at 640 connections for GET and 1,280 for SET; Redis and Valkey peak at 1,280. Redis uses 16 I/O threads; Valkey uses 16 for GET and 8 for SET; Lavik uses 16 workers. Each p99 corresponds to the selected throughput peak.

Tested Lavik commit: 3955b98d43b312324aa8d52775df52cfb111c0d0