Check out the new USENIX Web site. next up previous
Next: File system design Up: File system limitations Previous: Idleness.

UFS performance limitations

Table 1: Comparison of file system features.
feature UFS Hummingbird
name space hierarchical name space flat name space
reference locality application manages directories application sends locality hints
file meta-data meta-data kept on disk; synchronous updates most meta-data in memory
preserve consistency
disk layout of files data in blocks, with separate i-node for meta-data file and meta-data stored contiguously
disk access could be as small as a block size cluster size (typically 32 or 64 KB)
interface buffers are passed; memory copies are necessary pointers are passed

Due to the proxy workload characteristics presented in the previous section, UFS has a number of features which are not needed, or could be greatly simplified, so that file system performance could be improved. Table 1 presents the features on which UFS and a desired caching web proxy file system should differ. We now discuss these features.

UFS files are collections of fixed-size blocks, typically 8 KB in size. When accessing a file, the disk delays are due to disk head positioning occurring when the file blocks are not stored contiguously on disk. UFS attempts to minimize the disk head positioning time by storing the file blocks contiguously and prefetching blocks when a file is accessed sequentially, and does a good job of this for small files. Thus, when the workload consists of mostly small files, the largest component of disk delays are due to the reference stream locality not corresponding with the on-disk layout. UFS attempts to reduce this delay by having the user place files into directories, and locates files in a directory on a group of contiguous disk blocks called cylinder groups. Thus, reference locality is tied to naming. Here, the responsibility for performance lies with the application or user, who must construct a hierarchy with directory locality that matches future usage patterns. In addition, to reduce file lookup times, the directory hierarchy must be well-balanced and any single directory should not have too many entries. Squid attempts to balance the directory hierarchy, but in the process distributes the reference stream across directories, thus destroying locality. Apache maps files from the same origin server into the same directory. For specifying locality by a web proxy, a more direct and low-overhead mechanism can be used.

Experience with Squid and Apache has shown that it is difficult for web proxies to use directory hierarchies to their advantage [11]. Deep pathnames mean long traversal times. Populated directories slow down lookup further because many legacy implementations still do a linear search for the filename through the directory contents. The hierarchical name space allows files to be organized by separating them across directories, but this is not needed by a proxy. What the proxy actually needs is a flat name space and the ability to specify storage locality.

UFS file meta-data is stored in the i-node, which is updated using synchronous disk writes to ensure meta-data integrity. A caching web proxy does not require file durability for correctness, so it is free to replace synchronous meta-data writes with asynchronous writes to improve the performance (which is done with soft updates [6]) and eliminate the need to maintain much of the meta-data associated with persistence.

Traditional file systems also force two architectural issues. First, the standard file system interface copies from kernel VM into the applications' address space. Second, the file system caches file blocks in its own buffer cache. Web proxies manage their own application-level VM caches to eliminate memory copies and use private information to facilitate more effective cache management. However, web documents cached at the application level are likely to also exist in the file system buffer cache, especially if recently accessed from disk. This multiple buffering reduces the effective size of the memory. A single unified cache solves the multiple buffering and configuration problems. Memory copy costs can be alleviated by passing data by reference rather than copying. Both of these can be done using a file system implemented by a library that accesses the raw disk partition. Another approach for alleviating multiple buffering is using memory-mapped files as done by Maltzahn et al. [11].

next up previous
Next: File system design Up: File system limitations Previous: Idleness.
Liddy Shriver 2001-05-01