Check out the new USENIX Web site. next up previous
Next: Enhancing Derby Up: An Application-Aware Data Storage Previous: Introduction


In our previous work, we investigated a new file system design called Derby [15,14] that used idle remote memory for both read and write traffic. Reads and writes occurred at remote memory speeds while providing the disk persistence necessary for database transaction-oriented storage. Derby assumes a small fraction of the workstations were equipped with uninterruptable power supplies (Workstations with UPS or WUPS). The system operates as follows. All active data resides in memory. Read requests that cannot be satisfied from the local cache use a dynamic address table lookup to find the idle machine that holds the data. The request is sent to a server process on the remote machine that returns the data from its memory. Write requests occur in a similar manner but also send the written/modified data to one or more WUPS machines. The data is held temporarily in WUPS memory until the server process asynchronously writes the data to disk and informs the WUPS that the newly written data can be removed. By using WUPS for short-term persistence and disks for long-term persistence, Derby achieves disk persistence at remote memory speeds.

The advantage of Derby and similar main memory storage systems [18,17,24] is the ability to achieve traditional disk persistence at memory-speeds. However, disk persistence comes at a price. Such systems require special purpose hardware such as NVRAM or uninterruptable power supplies. In Derby, disk persistence increases the communication overhead between clients, servers, and WUPS servers. In addition, the disk system poses a potential bottleneck because all write traffic (including small, frequent operations such as file creation, deletion, and other metadata changes) hits the disk. Note that past file system analysis shows average file lifetimes to be quite short (80% of files and 50% of bytes last less than 10 minutes [4]). Thus, many files are likely to be written to cache and disk, read from the cache, and deleted, without ever being read directly from disk. This unnecessarily consumes CPU, bus, memory, disk bandwidth, and network resources in a distributed file system.

next up previous
Next: Enhancing Derby Up: An Application-Aware Data Storage Previous: Introduction
Todd Anderson