``kawipiko`` is a simple static HTTP server written in Go_, whose main purpose is to serve static content as fast as possible.
However "simple" doesn't imply "dumb" or "limited", instead it implies "efficiency" and removal of superfluous features, inline with UNIX's philosophy of `do one thing and do it well <https://en.wikipedia.org/wiki/Unix_philosophy#Do_One_Thing_and_Do_It_Well>`__.
As such ``kawipiko`` basically supports only ``GET`` (and ``HEAD``) requests and does not provide features like dynamic content, authentication, reverse proxying, etc.
However, ``kawipiko`` does provide something unique, that no other HTTP server offers: the static content is served from a CDB_ database with almost zero latency.
Moreover, the static content can be compressed (with either ``gzip`` or ``brotli``) ahead of time, thus reducing not only CPU but also bandwith and latency.
CDB_ databases are binary files that provide efficient read-only key-value lookup tables, initially used in some DNS and SMTP servers, mainly for their low overhead lookup operations, zero locking in multi-threaded / multi-process scenarios, and "atomic" multi-record updates.
This also makes them suitable for low-latency static content serving over HTTP, which this project provides.
With regard to performance, as described in the `benchmarks section <#benchmarks>`__, ``kawipiko`` is on par with NGinx, sustaining 72k requests / second with 0.4ms latency for 99% of the requests even on my 6 years old laptop.
However the main advantage over NGinx is not raw performance, but deployment and configuration simplicity, plus efficient management and storage of large collections of many small static files.
* (optionally) in order to reduce the serving latency even further, one can preload the entire CDB database in memory, or alternatively mapping it in memory (mmap_); this trades memory for CPU;
* "atomic" site content changes; because the entire site content is held in a single CDB database file, and because the file replacement is atomically achieved via the ``rename`` syscall (or the ``mv`` tool), all the site's resources are "changed" at the same time;
*``_wildcard.*`` files (where ``.*`` are the regular extensions like ``.txt``, ``.html``, etc.) which will be used if an actual resource is not found under that folder; (these files respect the hierarchical tree structure, i.e. "deeper" ones override the ones closer to "root";)
* support for custom HTTP response headers (for specific files, for specific folders, etc.); (currently only ``Content-Type``, ``Content-Length``, ``Content-Encoding`` and optionally ``ETag`` is included; additionally ``Cache-Control: public, immutable, max-age=3600`` and a few security related headers are also included;)
* (won't fix) the CDB database **maximum size is 4 GiB**; (however if you have a site this large, you are probabbly doing something extreemly wrong, as large files should be offloaded to something like AWS S3 and served through a CDN like CloudFlare or AWS CloudFront;)
* (won't fix) the server **does not support per-request decompression / recompression**; this implies that if the site content was saved in the CDB database with compression (say ``gzip``), the server will serve all resources compressed (i.e. ``Content-Encoding: gzip``), regardless of what the browser accepts (i.e. ``Accept-Encoding: gzip``); the same applies for uncompressed content; (however always using ``gzip`` compression is safe enough as it is implemented in virtually all browsers and HTTP clients out there;)
* (won't fix) regarding the "atomic" site changes, there is a small time window in which a client that has fetched an "old" version of a resource (say an HTML page), but which has not yet fetched the required resources (say the CSS or JS files), and the CDB database was swapped, it will consequently fetch the "new" version of these required resources; however due to the low latency serving, this time window is extreemly small; (**this is not a limitation of this HTTP server, but a limitation of the way the "web" is built;** always use fingerprints in your resources URL, and perhaps always include the current and previous version on each deploy;)
* under normal conditions (16 concurrent connections), you get around 72k requests / second, at about 0.4ms latency for 99% of the requests;
* under stress conditions (512 concurrent connections), you get arround 74k requests / second, at about 15ms latency for 99% of the requests;
***under extreme conditions (2048 concurrent connections), you get arround 74k requests / second, at about 500ms latency for 99% of the requests (meanwhile the average is 50ms);**
* (the timeout errors are due to the fact that ``wrk`` is configured to timeout after only 1 second of waiting;)
* (the read errors are due to the fact that the server closes a keep-alive connection after serving 256k requests;)
***the raw performance is comparable with NGinx_** (only 20% few requests / second for this "synthetic" benchmark); however for a "real" scenario (i.e. thousand of small files accessed in a random pattern) I think they are on-par; (not to mention how simple it is to configure and deploy ``kawipiko`` as compared to NGinx;)
* (the NGinx configuration file can be found in the `examples folder <./examples>`__; the configuration was obtained after many experiments to squeeze out of NGinx as much performance as possible, given the targeted use-case, namely many small static files;)
* 512 concurrent connections (handled by 2 threads): ::
wrk \
--threads 2 \
--connections 512 \
--timeout 6s \
--duration 30s \
--latency \
http://127.0.0.1:8080/ \
#
* 4096 concurrent connections (handled by 4 threads): ::
wrk \
--threads 4 \
--connections 4096 \
--timeout 6s \
--duration 30s \
--latency \
http://127.0.0.1:8080/ \
#
Methodology notes
.................
* the number of threads for the server plus for ``wkr`` shouldn't be larger than the number of available cores; (or use different machines for the server and the client;)
* also take into account that by default the number of "file descriptors" on most UNIX/Linux machines is 1024, therefore if you want to try with more connections than 1000, you need to raise this limit; (see bellow;)
* additionally, you can try to pin the server and ``wrk`` to specific cores, increase various priorities (scheduling, IO, etc.); (given that Intel processors have HyperThreading which appear to the OS as individual cores, you should make sure that you pin each process on cores part of the same physical processor / core;)
* pinning the server (cores ``0`` and ``1`` are mapped on physical core ``1``): ::
sudo -u root -n -E -P -- \
\
taskset -c 0,1 \
nice -n -19 -- \
ionice -c 2 -n 0 -- \
chrt -r 10 \
prlimit -n16384 -- \
\
sudo -u "${USER}" -n -E -P -- \
\
kawipiko-server \
... \
#
* pinning the client (cores ``2`` and ``3`` are mapped on physical core ``2``): ::
For details about the copyright and licensing, please consult the `notice <./documentation/licensing/notice.txt>`__ file in the `documentation/licensing <./documentation/licensing>`__ folder.
*`Benchmarking LevelDB vs. RocksDB vs. HyperLevelDB vs. LMDB Performance for InfluxDB <https://www.influxdata.com/blog/benchmarking-leveldb-vs-rocksdb-vs-hyperleveldb-vs-lmdb-performance-for-influxdb/>`__ (article);
*`Badger vs LMDB vs BoltDB: Benchmarking key-value databases in Go <https://blog.dgraph.io/post/badger-lmdb-boltdb/>`__ (article);
*`Benchmarking BDB, CDB and Tokyo Cabinet on large datasets <https://www.dmo.ca/blog/benchmarking-hash-databases-on-large-data/>`__ (article);
