Back in December when I was working on the Places branch, I was using Xperf quite a bit to try and figure out why we were regressing certain performance tests. Xperf is an incredibly powerful tool, but it’s really hard to get it to do what you want sometimes. This is largely due to the fact that the documentation isn’t great, and there appears to be wrong information on the Internet (from blogs, which tend to be more useful than the documentation).
I cared about getting information about hard faults, file I/O, and disk I/O to see if patterns changed with my work. In order to accomplish this, I started Xpef like this (from an admin console):
xperf -start "NT Kernel Logger" -on PROC_THREAD+LOADER+HARD_FAULTS+FILE_IO+FILE_IO_INIT+DISK_IO -stackwalk FileWrite+FileRead+FileFlush -MaxBuffers 1024 -BufferSize 1024 -f output.etl
The first list options after
-on is a list of providers. To see the list of installed kernel providers on your system and what they do, open up a command prompt, and run
xperf -providers K.
The options after
-stackwalk tells Xperf to get call stacks for certain events. The full list of events supported can be found here. I found it very difficult to get stacks, and if you are on a 64-bit version of windows, you have an extra hoop to jump through. I was actually never able to get stacks from an optimized build with symbols, so I ended up just using debug builds when I needed stacks, and opt builds when I wanted good numbers for everything else.
-BufferSize arguments were useful to prevent events from being dropped (if Xperf doesn’t have enough memory set aside to record an event, it just drops it). You can tweak those values to your needs, but the values I used should be fine.
Once I ran that command in my console, I’d perform the test I wanted to get data on. Upon completion, Xperf needs to be told to stop, and then have it merge (merging may be unnecessary when you only use kernel providers, but I never tested this theory):
xperf -stop "NT Kernel Logger"
xperf -merge output.etl output_final.etl
You can now open
output_final.etl to examine the data you just gathered!