SAFS design
SAFS splits a file into multiple partitions and each SSD gets a partition of the file. SAFS deploys a Linux filesystem on the SSDs and each partition is stored as a native file in the Linux filesystem. To map data in a SAFS file to SSDs, SAFS first splits the file into blocks of the user-defined size and map each block to a particular SSD and a particular location in the native file on the SSD with a map function. Therefore, SAFS does not need to maintain metadata to get the physical location of a data block in an SAFS file. There are three map functions available right now:
- Striping: Data are divided into fixed-size small blocks placed on successive disks in increasing order.
- Rotated Striping: Data are divided into stripes but the start disk for each stripe is rotated, much like distributed parity in RAID5.
- Hash mapping: The placement of each block is randomized among all disks.
The figure below shows the data layout of an SAFS file. In this example, the SAFS file has 9 blocks and each SSD gets 3 blocks. The 3 blocks on an SSD are stored in a native file. This figure also illustrates the Striping map function. The first data block in the SAFS file is mapped to the first data block in the native file on the first SSD, and the second data block in the SAFS file is mapped to the first data block in the native file on the second SSD, and so on.
The details of SAFS design is described in the paper published in SuperComputing'13.