flashmem

A high-performance in-memory store written in C++.

Architecture

Shared-nothing, lock-free architecture. Data is partitioned across multiple threads. Any thread can handle a request depending on the underlying OS. On Linux, the kernel will load balance a TCP port over multiple threads. On other OS, only one thread will establish the connection, but will pass data across to the correct thread depending on a request key.

Flashmem Protocol (FMP)

Uses a custom binary protocol over TCP called FMP. Features implemented:

• Pipelining: avoids head-of-line blocking
• Persistent connections: multiple requests per connection
• Python client

Request Structure

A request consists of a fixed-size header followed by a variable number of "parts" (arguments).

Section	Field	Size	Description
Header	Magic Number	4 bytes	0xCAFEBEEF (Sanity check)
Header	Total Length	4 bytes	Total size (Header + Body)
Header	Request ID	4 bytes	Unique ID
Body	Part Length	4 bytes	Length of the argument
Body	Part Data	Variable	The argument data

Response Structure

Section	Field	Size	Description
Header	Magic Number	4 bytes	0xCAFEBEEF
Header	Total Length	4 bytes	Total size
Header	Request ID	4 bytes	Matches request ID
Header	Status	4 bytes	Status code
Body	Payload	Variable	Response data

Roadmap

• Client
  • Connection pooling: one client should be able to handle many concurrent requests. request_id is already specified in FMP.
  • Messaging multiplexing
• Server
  • Messaging multiplexing
  • More stack usage to maximize performance