This project is a fork of Redis, adapted for systems with multiple memory tiers. New memory technologies include memories that have very high bandwidth at the cost of latency, slower large capacity at lower cost, sharing memory among multiple machines in a rack, and so on.
The overall performance of a system can be improved if frequently accessed ("hot") data is kept in a faster tier while "warm" data that still needs to be readily available in memory (rather than disk or similar slow storage) yet is not as critical, may be kept in a slower memory tier.
This README doesn't describe usual usage of Redis. We assume you already have general experience with Redis.
Generally, TieredMemDB is close to vanilla Redis, and can be used on the same set of systems. You can even build without tier support, as all functionality of Redis should be left intact. To actually get the benefits of memory tiers, you need the following extra dependencies:
* ndctl and daxctl
* memkind (included in tarballs)
Building TieredMemDB is as simple as:
% make
As usual, you can -- and probably want to -- confirm the built code works well:
% make test
The crux of TieredMemDB is the tiered allocator, which beside MALLOC=libc
and MALLOC=jemalloc supported by Redis adds:
% make MALLOC=memkind
but in release tarballs this is already the default.
The set of patches over base Redis that comprises TieredMemDB can be applied onto your modified versions of Redis as well. To do that, or to update Redis to a new minor version (because of eg. a security update), you rebase the tree available on GitHub, or ask git to produce a patch, with our branch checked out:
% git diff 6.2.6 @
which can then be applied via patch over, in this case, Redis 6.2.6.