CORFU organizes a cluster of flash devices as a single, shared log that can be accessed concurrently by multiple clients over the network. [It] can be viewed as a distributed SSD
A shared log over a cluster of flash units that many clients append to and read from at once, with strong consistency (linearizable).
Each position in the shared log is mapped to a set of flash pages on different flash units. This map is maintained – consistently and compactly – at the clients.
Components
- A mapping function from logical positions in the log to flash pages on the cluster of flash units.
- A tail-finding mechanism for finding the next available logical position on the log for new data.
- A replication protocol to write a log entry consistently on multiple flash pages.
To read data at a specific log position, the client-side library uses the mapping function to find the appropriate flash page, and then directly issues a read to the device where the flash page is located.
To append data, a client finds the tail position of the log, maps it to a set of flash pages, and then initiates the replication protocol that issues writes to the appropriate devices
Flash Unit Requirements
- Reads on pages that have not yet been written should return an error code (
error_unwritten) - Writes on pages that have already been written should also return an error code (
error_overwritten) - Required to support a trim command, allowing clients to indicate that the flash page is not in use anymore
- Required to support a seal command. Each incoming message to a flash unit is tagged with an epoch number. When a particular epoch number is sealed at a flash unit, it must reject all subsequent messages sent with an epoch equal or lower to the sealed epoch.
- Expected to detect and re-map bad blocks in this address space
Handling Failures
- A hole is a log position obtained from a sequencer for an append where the client crashes without completing the write to that position
- This is handled with a fast hole-filling primitive