Disk Workloads

There are six different basic access patterns. Read, write, and update operations can either be sequential or randomly distributed. Sequential read and write occur when files are copied and created or when large amounts of data are being processed. Random read and write can occur in indexed database reads or can be due to page-in or page-out to a file. Update consists of a read-modify-write sequence and can be caused by a database system committing a sequence of transactions in either a sequential or random pattern. When you are working to understand or improve the performance of your disk subsystem, spend some time working out which of these categories you expect to be most important.

You cannot automatically tell which processes are causing disk activity; the kernel does not collect this information. You may be able to work out where the workload comes from by looking at how an application was installed, but often you must resort to using truss on a process or the TNF tracing system. See “Tracing Applications” on page 155. The use of TNF to trace disk activity is covered in “The Solaris 2.5 Trace Capability” on page 188. I would like to see a way of getting I/O activity on a per-file descriptor basis added to Solaris, but until that happens, application-specific instrumentation is all you have. Databases such as Oracle can collect and report data on a per-tablespace basis, so if you can map the tablespaces to physical disks, you can tell what is going on programmatically. This kind of collection and analysis is performed by the BGS Best/1 performance modeling tool, so that changes in disk workload can be modeled.