Skip to content

Latest commit

 

History

History
80 lines (56 loc) · 4.15 KB

README.md

File metadata and controls

80 lines (56 loc) · 4.15 KB
Raw

SWARM-KV

SWARM is a wait-free replication protocol for shared data in disaggregated memory that provides 1-RTT reads and writes.

SWARM-KV is an RDMA-based disaggregated key-value store that leverages SWARM to offer 1-RTT GETs and UPDATEs.

Requirements

  • conan package manager

    pip3 install --user "conan>=1.47.0,<2.0"

    make sure to set the default ABI to C++11 with:

    conan profile new default --detect  # Generates default profile detecting GCC and sets old ABI
conan profile update settings.compiler.libcxx=libstdc++11 default  # Sets libcxx to C++11 ABI

  • cmake v3.9.x

  • clang-format >= v6.0.0

Build

Run from within the root:

./build.py swarm-kv

This will create all conan packages required by swarm-kv, and its executables.

Note: If gcc is available, it is used as the default compiler. In a system with clang only, then clang becomes the default compiler. In any case, you can check the available compilers/compiler versions by calling ./build.py --help.

Docker

You can manually build the Dockerfile under the root of this repo.

docker build -t dory .

Usage

Refer to SWARM-KV's artifact repository for detailed instructions on how to run the experiments.

Navigating the code

Unfortunately, the algorithms presented in the paper were heavily factorized vs our implementation, so there's no clearly defined place where abstraction X or Y is implemented, but the rough mapping is:

  • In-n-Out and the CAS-based MAX-replacement are located in unreliable_maxreg.hpp.
  • relaxed_maxreg.hpp uses the above and provides the replicated max register abstractions and ensures operations are executed on a majority of servers.
  • oops_data_future.hpp uses the above and provides Safe-Guess' logic.
  • The above components implements state machines where the "step" is the current state in the respective algorithms and RDMA requests are sent by unreliable_maxreg.hpp when transitioning between steps.
  • Timestamp validators are part of the RTryNotifyWriter and WTryRepare steps of oops_data_future.hpp.
  • tspointers.hpp corresponds to metadata buffers that both contain a timestamp and a pointer to the out-of-place data.
  • oops_client.hpp initiates and manages all instances of oops_data_future.hpp.
  • oops_state.hpp stores state and configuration shared between the client and all the underlying futures.
  • layout.hpp contains the system parametters and allow for easy translations from element indices to RDMA-accessible memory locations in clients and servers.
  • latency.hpp allows for efficient percentile latency measurements.
  • main.cpp handles all the arguments, bootstraps the client or server, and generates and executes the workloads on clients.
  • The race subfolder contains a simplified implementation of RACE, a low-latency RDMA-based index.

All the above are contained in the swarm-kv submodule.

The repository also includes the following submodules:

  • fusee: an implementation of the FUSEE KVS
  • conn: provides interfaces to easily create and manage reliable connections over RDMA.
  • ctrl: provides interfaces to manage RDMA devices and control blocks.
  • memory: provides datastructures used by ctrl
  • memstore: provides an interface to access memcached. This module is used by conn to simplify the coordination of clients and servers for the configuration of RDMA connections.
  • special: provides cmd-lines tools
  • extern: provides the ibverbs and memcached external librairies
  • shared: provides various tools used by multiple submodules
  • third-party: provides third-party tools (used by conn)

The dependencies are described in targets.yaml