1The following are additional notes on the iosnoop program, 2 3 4* What does the output represent? 5 6The output is disk events - I/O operations that cause the disk to physically 7read or write the data. 8 9The output is not application I/O events that are absorbed by the cache - 10many of which will be. The output really is physical disk events. 11 12 13 14* What do the elapsed and delta times mean? 15 16Glad you asked! 17 18The times may *not* be as useful as they appear. I should also add that 19this quickly becomes a very complex topic, 20 21There are two different delta times reported. -D prints the 22elapsed time from the disk request (strategy) to the disk completion 23iodone); -o prints the time for the disk to complete that event 24since it's last event (time between iodones, or since idle->strategy). 25 26The elapsed time is equivalent to the response time from the application 27request to the application completion. The delta time resembles the 28service time for this request (resembles means it will be generally 29correct, but not 100% accurate). The service time is the the time for the 30disk to complete the request, after it has travelled through any bus or 31queue. 32 33buuuttt.... you need to think carefully about what these times mean before 34jumping to conclusions. For example, 35 36 You troubleshoot an application by running iosnoop and filtering 37 on your application's PID. You notice large times for the disk events 38 (responce, service, for this example it doesn't matter). 39 Does this mean there is a problem with that application? 40 What could be happening is that a different application is also using 41 the disks at the same time, and is causing the disk heads to seek to 42 elsewhere on the disk surface - increasing both service and response time. 43 44hmmm! so you can't just look at one application, one set of numbers, and 45understand fully what is going on. 46 47But it gets worse. Disks implement "tagged queueing", where events in the 48queue are reshuffeled to promote "elevator seeking" of the disk heads (this 49reduces head seeking). So the time for a disk event can be effected not 50just by the previous event (and previous location the heads had seeked to), 51but the surrounding events that enter the queue. 52 53So the good and the bad. The good news is that iosnoop makes it easy to 54fetch disk event data on a live system, the bad news is that understanding 55all the data is not really easy. 56 57For further information on disk measurements see, 58 59 "How do disks really work?" - Adrian Cockcroft, SunWorld Online, June 1996 60 "Sun Performance and Tuning" - Adrian Cockcroft, Richard Pettit 61 "Solaris Internals" - Richard McDougall, Jim Mauro 62 63 64 65* The output appears shuffled? 66 67Read the answer to this in ALLsnoop_notes.txt. 68 69