reading ski trace files

Just wondering what tool one is meant to use to read the traces produced by ski? A hexdump is vaguely helpful, but the file format suggests that it's intended to be read by some automated tool :)

Thanks,
Matt

Matt> Just wondering what tool one is meant to use to read the traces Matt> produced by ski? A hexdump is vaguely helpful, but the file format Matt> suggests that it's intended to be read by some automated tool :)

Yes, there is a tool for doing that, but it hasn't been released yet. I need to check with the author to see if it would be OK to do so. Ping me if you don't here back some time after next week.

	--david

Matt> Just wondering what tool one is meant to use to read the Matt> traces produced by ski? A hexdump is vaguely helpful, but the Matt> file format suggests that it's intended to be read by some Matt> automated tool :)

OK, you can now get the following file:

	ftp://ftp.hpl.hp.com/pub/linux-ia64/trcread-x86-020329.tar.gz

If you extract it, you'll get an x86 binary called "trcread". This program will let you print the trace files collected by ski. The trace file contains tons of information and the output of "trcread" is not really easy to interpret (partly due to lack of documentation). However, if you just want to generate a memory trace, I'd recommend to use the -d option. Interspersed with the normal output, this will output lines of the form:

mem:FOO

There is one such line for each memory access in the trace. FOO is a line describing the memory access in a format compatible with the Dinero IV cache simulator. Briefly, this format is as follows:

mem:i 00000000001014e0 10

	=> read 16 bytes of instruction from address 00000000001014e0
 mem:r 00000000001035b0 8
	=> read 8 bytes of data from address 00000000001035b0
 mem:w 0000000000103688 4
	=> write 4 bytes of data to address 0000000000103688

mem:m e0000000004dfc20 1
=> prefetch cache line from address e0000000004dfc20 (the 1 byte
"size" is meaningless)

Semaphore instructions (fetchadd etc.) are output like regular stores (i.e., there is no separate record for the implied atomic read).

As an example, to get a memory trace with ski, you could something along the lines of:

	$ ski arch/ia64/boot/bootloader vmlinux single
	* trfile /tmp/foo.trace		# set trace file
	* trace 1			# turn on tracing
	* run				# run time program for a while
	* trace 0			# turn tracing off

Then you could get the Dinero records with:

	$ trcread -d -i /tmp/foo.trace | grep '^mem:'

Needless to say, this program is provided as is, with no warranty, etc, etc. But if you find a bug or have a question, please do send mail to [email protected].

Thanks,

	--david