Alan Hargreaves' Weblog

DTrace Lab in the performance track at CEC

We have just been helping out in the DTrace lab for the performance track at CEC. On having a look at exercise 1, we suggested to the attendees that they attempt the other exercises and come back to this one. The exercise was:

Enhance /usr/demo/dtrace/iosnoop to be able to detect reads that are satisfied by the filesystem cache (UFS: pagecache).

We then went off to write our own versions of an answer to the problem.I started from scratch, and (referring back to a previous blog where I did some digging into this code), came up with the following:

#!/usr/sbin/dtrace -s

#pragma D option quiet

BEGIN { printf("Collecting ...\n"); }

syscall::read*:entry {
        self->interest = 1;
        self->file = fds[arg0].fi_name;
        self->phys = 0;
}

/* we only increment this kstat if we have to go to disk, from a filesystem */

sysinfo:::bread /self->interest/ { self->phys = 1;}

syscall::read*:return /self->interest/ {
        this->str = self->phys
            ? "Physical reads"
            : "Cache Reads";
        @[this->str, self->file] = count();
        self->interest = 0;
}

END { printa("%@6d %s %s\n", @); }

On running this you get output like:

   ...
   209 Cache Reads jaxrpc-impl.jar
   211 Cache Reads appserv-admin.jar
   311 Cache Reads xalan.jar
   354 Cache Reads ttysrch
   372 Cache Reads jaxr-impl.jar
   373 Cache Reads auxv
   500 Cache Reads appserv-assemblytool.jar
  1216 Cache Reads 
  1324 Cache Reads appserv-rt.jar
  1392 Cache Reads clone@0:ptm
 83181 Cache Reads psinfo

Of course by the time I finished writing this, pretty much everything in the only ufs filesystem on the box had been cached :)

Posted at 02:01PM Nov 14, 2008 by Alan Hargreaves in Solaris  |