ReleaseNotes150
Date: 2017-7-21
http://www.memcached.org/files/memcached-1.5.0.tar.gz
This release makes the options -o modern enabled in 1.4.39 the default
options. It also adds --long-options (see --help|-h) if the platform supports
it and fixes a small bug.
People who were not running -o modern before should see wins in hit rate or reductions in memory requirements. Please report any regressions.
- fix for musl libc: avoid huge stack allocation
This release is the culmination of many years of part-time work.
Memcached was introduced in 2003, two years before the first dual-core opteron x86 processor. Since then computers have added huge numbers of cores and vastly expanded memory.
Recently the cost ratio of memory to CPU has started to shift a bit. For a few years memory prices will likely be inflated (though this always comes in cycles). I've managed to keep memcached's performance and thread scalability very high. Now it makes use of some extra CPU in background threads in order to reduce memory usage. You may find you need half as much RAM to get the same hit ratio, or it could be the same as what you have now.
The new features:
- LRU crawler to background-reclaim memory. Mixed-TTL's and LRU reordering leaves many holes, making it difficult to properly size an instance.
- Segmented LRU. HOT/WARM/COLD and background processing should try harder to keep semi-active items in memory for longer.
- Automated slab rebalancing. Avoiding slab stagnation as objects change size over time.
- Faster hash table lookups with murmur3 algorithm (though it's been so long this is now outdated again;)
- Reduce memory requirements per-item by a few bytes here and there
- Immediately close connections when hitting the connection limit, instead of hanging until a spot opens up.
- Items larger than 512k (by default) are assembled by stacking multiple chunks together. Now raising the item size above 1m doesn't drop memory efficiency by spreading out slab classes.
These features have been slowly assembled, tuned, and tested behind feature flags for far too long. Recently after finishing reworking the automated page mover, I felt confident enough to flip the defaults around and call this thing 1.5.x.
-o modern will live on as new feature flags appear. Though they may not take
as long to migrate to the actual defaults.
If you want the old defaults, -o no_modern will work for a while.
Some old flags will be deprecated (old maxconns handling, and the inverse of the byte-savings, for two) and eventually removed.
A great thanks goes out to Andrew Rodland (hobbs) for running production tests of the adjusted slab page algorithm. Help from the community greatly reduces the time it takes for me to find bugs and gain confidence in features.
Another great thanks to https://www.packet.net/ - Their donation of test hardware has been invaluable while I run endless burn-in tests on various features. It's crazy the weird bugs you find while shoveling tens of millions of qps at a single instance while moving memory around randomly!
It'd be great to hear from users as well. It's been a long time and there are a lot of key/value stores now. If this work is truly useful we'll keep it up.
The following people contributed to this release since 1.4.39.
Note that this is based on who contributed changes, not how they were done. In many cases, a code snippet on the mailing list or a bug report ended up as a commit with your name on it.
Note that this is just a summary of how many changes each person made
which doesn't necessarily reflect how significant each change was.
For details on what led up into a branch, either grab the git repo and
look at the output of git log 1.4.39..1.5.0 or use a web view.
- Repo list: https://github.com/memcached/memcached/wiki/DevelopmentRepos
- Web View: http://github.com/memcached/memcached/commits/1.5.0
2 dormando