File System Logging versus Clustering: A Performance Comparison

The Log-structured File System (LFS), introduced in 1991 [8], has received much attention for its potential order-of-magnitude improvement in file system performance. Early research results [9] showed that small file performance could scale with processor speed and that cleaning costs could be kept low, allowing LFS to write at an effective bandwidth of 62 to 83% of the maximum. Later work showed that the presence of synchronous disk operations could degrade performance by as much as 62% and that cleaning overhead could become prohibitive in transaction processing workloads, reducing performance by as much as 40% [10]. The same work showed that the addition of clustered reads and writes in the Berkeley Fast File System [6] (FFS) made it competitive with LFS in large-file handling and software development environments as approximated by the Andrew benchmark [4].

These seemingly inconsistent results have caused confusion in the file system research community. This paper presents a detailed performance comparison of the 4.4BSD Log-structured File System and the 4.4BSD Fast File System. Ignoring cleaner overhead, our results show that the order-of-magnitude improvement in performance claimed for LFS applies only to meta-data intensive activities, specifically the creation of files one-kilobyte or less and deletion of files 64 kilobytes or less.

For small files, both systems provide comparable read performance, but LFS offers superior performance on writes. For large files (one megabyte and larger), the performance of the two file systems is comparable. When FFS is tuned for writing, its large-file write performance is approximately 15% better than LFS, but its read performance is 25% worse. When FFS is optimized for reading, its large-file read and write performance is comparable to LFS.

Both LFS and FFS can suffer performance degradation, due to cleaning and disk fragmentation respectively. We find that active FFS file systems function at approximately 85-95% of their maximum performance after two to three years. We examine LFS cleaner performance in a transaction processing environment and find that cleaner overhead reduces LFS performance by more than 33% when the disk is 50% full.