Sleipnir v1.1.0

Speed Improvements

Sleipnir v1.1.0 is a performance-focused release. The engine that serves your sites now does noticeably less work per request and handles heavy concurrency more evenly across workers while staying the same lean, dependency-free binary (~258 KB).

What changed

• Fewer system calls per request. On the static hot path we cut the work the kernel has to do from roughly 16 syscalls per request down to about 6 - caching directory .htaccess look-ups, sending small files in a single write, and trimming redundant timer and epoll calls.
• Even connection handling across workers. New connections now wake a single worker instead of all of them (EPOLLEXCLUSIVE), so throughput scales with workers under heavy connection churn instead of degrading.
• Same features, same footprint. No change to behaviour or configuration, just faster.

⚠ How these were tested. Every web server below was benchmarked directly, with no proxy in front of it, so the comparison is apples-to-apples between the servers themselves. In production, Unicorn runs Sleipnir behind a reverse proxy (which handles TLS, caching and routing), so real-world performance reflects that additional layer - these numbers isolate the web server, not the full stack.

Static files: requests/sec (keep-alive)

Higher is better. Columns are concurrent clients: light, heavy, and extreme load.

* Apache returned connection errors (dropped/reset connections) at high concurrency; Sleipnir, nginx and OpenLiteSpeed served all three levels cleanly with zero errors, including 10,000 simultaneous clients.

PHP: requests/sec (50 concurrent, keep-alive)

Sleipnir, nginx and Apache all ran the same PHP-FPM pool, so that is a like-for-like comparison. OpenLiteSpeed uses its own LSAPI PHP engine, so its PHP figures reflect a different backend and are not directly comparable.

Takeaway

On the workloads that matter for real sites — static assets and PHP/WordPress — Sleipnir 1.1.0 is at or near the top of the field on static files, level with nginx on PHP, and well ahead of Apache, serving up to 10,000 concurrent clients without a single dropped connection.

Testing notes

• All four servers ran on the same 8-core Linux VM, over localhost, with equal worker counts and access logging disabled on every server, serving identical files.
• These are lab figures meant to compare the servers fairly; they are not a controlled datacenter benchmark, and run-to-run variance applies - hence the rounded numbers.
• As noted above, servers were tested directly; Unicorn's production stack adds a reverse proxy in front.