Structural changes, such as file creation and block allocation, have consistently been identified as file system performance problems in many user environments. We compare several implementations that maintain metadata integrity in the event of a system failure but do not require changes to the on-disk structures. In one set of schemes, the file system uses asynchronous writes and passes ordering requirements to the disk scheduler. These schedulerenforced ordering schemes outperform the conventional approach (synchronous writes) by more than 30 percent for metadata update intensive benchmarks, but are suboptimal mainly due to their inability to safely use delayed writes when ordering is required. We therefore introduce soft updates, an implementation that asymptotically approaches memory-based file system performance (within 5 percent) while providing stronger integrity and security guarantees than most UNIX file systems. For metadata update intensive benchmarks, this improves performance by more than a factor of two when compared to the conventional approach. 1

This dissertation demonstrates that the conventional approach for evaluating the performance of an I/O subsystem design, which is based on standalone subsystem models, is too narrow in scope. In particular, conventional methodology treats all I/O requests equally, ignoring differences in how individual request response times affect system behavior. As a result, it often leads to inaccurate performance predictions and can thereby lead to incorrect conclusions and poor design choices. A new methodology, which expands the model's scope to include other important system components (e.g., CPUs and system software), is proposed and shown to enable accurate predictions of both subsystem and overall system performance. This dissertation focuses on two specific problems with conventional methodology: 1. Benchmark workloads are often not representative of reality in that they do not accurately reflect feedback effects between I/O subsystem performance (in particular, individual request completion times) and the workload of requests (in particular, subsequent request arrivals). 2. Changes in I/O subsystem performance (e.g., as measured by mean request response times) do not always translate into similar changes in overall system performance (e.g., as measured by mean elapsed times for user tasks). These problems are fundamental to the subsystem-oriented approach and are independent of the model's accuracy. The first problem is illustrated with several examples where commonlyutilized workload generators trivialize feedback effects and produce unrealistic workloads. In each case, quantitative and/or qualitative errors result. The second problem is illustrated with a disk sche...

Structural changes, such as file creation and block allocation, have consistently been identified as file system performance problems in many user environments. We compare several implementations that maintain metadata integrity in the event of a system failure but do not require changes to the on-disk structures. In one set of schemes, the file system uses asynchronous writes and passes ordering requirements to the disk scheduler. These schedulerenforced ordering schemes outperform the conventional approach (synchronous writes) by more than 30 percent for metadata update intensive benchmarks, but are suboptimal mainly due to their inability to safely use delayed writes when ordering is required. We therefore introduce soft updates, an implementation that asymptotically approaches memory-based file system performance (within 5 percent) while providing stronger integrity and security guarantees than most UNIX file systems. For metadata update intensive benchmarks, this improves perform...