CORFU Sample Clauses

CORFU. ‌ In trying to search for an effective layering model, we sought inspiration in a system called CORFU [39]. It defines the abstraction of a single log shared across multiple devices, implemented over a cluster of flash storage that can be accessed by multiple clients concurrently over the network. The log works as a map, linking each map position to a flash storage unit. XXXXX’s main goal is to provide a simplified and inexpensive interface for running applications to interact with the underlying flash storage. With this in mind, CORFU is implemented as a client-side library that defines an Application Programming Interface (API) with the following main methods. append adds data to the log and returns its position in the log; read returns the data in the given position; trim removes out of date/invalid data from the log, at the request of the application; fill inserts ”junk” into a certain log position to prevent it from being updated in the future. This interface provides a clear separation and a high degree of independence between the two layers. Therefore, the shared log logic is simplified to the point of being faster and of easy and concurrent access, while the client logic can be as complex as necessary. Implementing the previous interface requires three functions. First, a deterministic mapping function that projects the positions on the log to the flash pages stored in the units. For example, if clients want to read from the log, they first consult their local map to check which flash units can be used and then apply the function to calculate the page to be used, reading from it directly. In the case of any flash unit crashing, the client is automatically remapped to a different flash unit. Second, a tail-finding mechanism: to append data, a client must first ask a sequencer node to assign them the next empty position (the tail) on the log, and only then they can write to it. Third, a chaining replication protocol: if the client wants to write to a group of units, it does so sequentially, only writing to the second one after getting confirmation on the first one, thus forming a chain. This protocol guarantees an important property: safety under contention, which guarantees that if two clients are writing at the same time in the chain, only one can append the value, while the other gets an error. Therefore, any subsequent read on that position will always return a single value: the one successfully appended – a property called single-copy semantics....