I describe small file clustering file system, a new mechanism for improving disk write performance in the environment in which there are many small files and they are accessed intensively.(e.g, compiling kernel, web proxy server, e-mail server, etc.) Traditional file systems have tried to exploit maximum disk bandwidth as much as possible. For example, FFS(Fast File System) is popular file system because of its good read and write performance. However, its block allocation mechanism is still not successful at accessing small files. Many small files scattered on the disk layout cause several overheads by separated I/O requests(e.g, rotational delay, seek time, computational overhead). The proposed mechanism, called small file clustering file system, is based on delayed block allocation mechanism of which key idea is that no need of allocating physical block number until write back time when dirty blocks are written to the disk by synchronization routine. This mechanism can figure out the logical locality of related blocks to be clustered at multi block allocation time and eliminate reallocation overhead for clustering. Experiment results in both sequential and random write operation of a large number of small files in many directories showed that small file clustering file system provides consistently a significant throughput improvement (6% to 30%) compared to previous FFS file system. I expect that small file clustering file system will be mostly useful for many applications such as web proxy server, e-mail server and Netnews server, which are the environments in which many small files are accessed intensively.
