There are a number of questions that are still unanswered by the first two benchmarks. First, the VLD always has plenty of free space in the previous benchmarks. How much degradation can we expect when the disk is fuller? Second, the LFS cleaner is not invoked so far. What is the impact of the cleaner under various disk utilizations? Third, we know that LFS performs poorly with frequent flushes to the disk. How much can NVRAM help? We attempt to answer these questions with the third benchmark. Again, our goal is not to compare UFS against LFS. Instead, our interest is in discovering how virtual logging can speed up synchronous writes on UFS and how NVRAM can improve LFS.
In this benchmark, we create a single file of a certain size. Then we repeatedly choose a random 4 KB block to update. There is no idle time between writes. For UFS, the ``write'' system call does not return until the block is written to the disk surface. For LFS, we assume that the 6.1 MB file buffer cache is made of NVRAM and we do not flush to disk until the buffer cache is full. We measure the steady state bandwidth of UFS on the regular disk, UFS on the VLD, and LFS on the regular disk as we vary the size of the file we update.
Figure 8 plots the average latency experienced per write. UFS on the regular disk suffers from excessive disk head movement due to the update-in-place policy. The latency increases slightly as the updated file grows because the disk head needs to travel a greater distance between successive writes. This increase may have been larger had we simulated the entire disk.
LFS provides excellent performance when the entire file fits in NVRAM. As soon as the file outgrows the NVRAM, writes are forced to the disk; as this happens and as disk utilization increases, the cleaner quickly dominates performance. The plateau between roughly 60% and 85% disk utilization is due to the fact that the LFS cleaner chooses less utilized segments to clean; with certain distribution patterns of free space, the number of segments to clean in order to generate the same amount of free segments may be the same as (or even larger than) that that required under a higher utilization.
With eager writing, the VLD suffers from neither the excessive disk head movements, nor the bandwidth waste during cleaning. As the disk utilization increases, the VLD latency also rises. The rise, however, is not significant compared to the various overheads in the system, which we examine next.