Sunday, June 3, 2012

Intel TurboBoost and Linux

My most recent rig, a Core i7-3930k, supports the very useful TurboBoost feature from Intel.  This essentially looks at the utilization of each core, bringing the lower utilized cores to a deep C-state when possible.  This effectively creates more thermal headroom to overclock the more active cores to higher frequencies (somewhere around 3.8 GHZ).

To do this, I had to configure my kernel (it is running Gentoo Linux, by the way) to allow for CPU frequency scaling and to utilize P-States.  I also had to enable the TurboBoost feature in the BIOS (not even sure that name is appropriate for modern era motherboards).

Doing all of this was fairly straightforward.  What wasn't straightforward, however, was grabbing the CPU clock frequency at any given interval.  Part of the problem is that the TurboBoost feature doesn't exactly correspond to clock frequency in the traditional ACPI sense.  The measured clock frequency through traditional procfs interfaces and cpu power tools will display the standard (i.e. stock) clock frequency (and below if you use the corresponding governor).  I wanted to -
A.) ensure that Turbo Boost was actually working and the performance gains I was experiencing weren't in fact placebos, and B.) I wanted a mechanism to share this information every 2.5 seconds in xmobar.  

In recent years there have been posts on LKML indicating a utility which is made for just this.  The utility is turbostat (and cpupower, to an extent).  Initial blog posts and LWN posts pointed that it was on gitorious but a more thorough search showed that the source code to these utilities is now being distributed with the kernel source tree.  The code can be found under tools/power/{x86/turbostat,cpupower}.  After compiling these utilities I was able to observe (unfortunately only as root), that my CPU both supported and was utilizing TurboBoost.

The output of turbostat looks a little something like this:


cor CPU    %c0  GHz  TSC    %c1    %c3    %c6    %c7   %pc2   %pc3   %pc6   %pc7
          0.36 2.44 3.20  23.14   0.52   0.00  75.97   0.00   0.00   0.00   0.00
  0   0   1.04 1.99 3.20   1.10   0.99   0.02  96.85   0.00   0.00   0.00   0.00
  0   6   0.12 2.31 3.20   2.01   0.99   0.02  96.85   0.00   0.00   0.00   0.00
  1   1   0.49 1.43 3.20   1.03   0.88   0.00  97.60   0.00   0.00   0.00   0.00
  1   7   0.14 1.70 3.20   1.38   0.88   0.00  97.60   0.00   0.00   0.00   0.00
  2   2   0.74 2.98 3.20  55.78   0.25   0.00  43.22   0.00   0.00   0.00   0.00
  2   8   0.07 1.82 3.20  56.46   0.25   0.00  43.22   0.00   0.00   0.00   0.00
  3   3   0.21 1.77 3.20   0.21   0.02   0.00  99.56   0.00   0.00   0.00   0.00
  3   9   0.06 2.17 3.20   0.35   0.02   0.00  99.56   0.00   0.00   0.00   0.00
  4   4   0.53 3.16 3.20  79.42   0.38   0.00  19.66   0.00   0.00   0.00   0.00
  4  10   0.57 3.56 3.20  79.38   0.38   0.00  19.66   0.00   0.00   0.00   0.00
  5   5   0.12 1.77 3.20   0.32   0.62   0.00  98.93   0.00   0.00   0.00   0.00
  5  11   0.18 2.20 3.20   0.27   0.62   0.00  98.93   0.00   0.00   0.00   0.00

Where it shows corresponding averages for the time spent in a given c-state.  Passing a -s parameter gave instead a summary of all of the cores (and only printed the column headers once).  There were a couple of problems with parsing this, however.

1.) You needed to be root to execute it.
2.) It did not self terminate, it needed a kill signal sent by a keyboard interrupt

While there may have been a way to fork an instance off and grab only two lines with bash, FIFOs, and shell scripts, I took the path of least resistant and wrote a quick C application meant to be run with setuid privileges.  This allowed me to open a pipe and close it off, effectively terminating the continuous output of turbostat (I'd rather have had it work more like iostat and friends where it allows for an argument to specify how many times to print).

Here is the source to my "turboinfo" program:

#include <stdio.h>
#include <stdlib.h>
#include <unistd.h>
#include <sys/types.h>

void usage(char *progname)
{
    fprintf(stderr,"Usage: %s [-i <interval>]\n",progname);
    exit(1);
}

int main(int argc, char *argv[])
{
    int arg;
    size_t alloc = 1000;
    uid_t userid = getuid();

    while ((arg = getopt(argc,argv,"")) != -1) {
        switch(arg) {
            default:
                usage(argv[0]);
                break;
        }
    }

    //const char *basecmd = "sudo turbostat -s -i 1 2>&1";
    char ghz[8];
    char *junkbuffer = malloc(sizeof(char)*1000);
    const char *basecmd = "turbostat -s -i 1 2>&1";

    seteuid(0);
    setuid(0);
    FILE *turboPipe = popen(basecmd,"r");
    seteuid(userid);
    setuid(userid);

    /* ignore first line */
    getline(&junkbuffer,&alloc,turboPipe);
    fscanf(turboPipe,"%*s %s %*s %*s %*s %*s %*s %*s %*s %*s %*s\n",ghz);
    free(junkbuffer);

    fprintf(stdout,"%s\n",ghz);
    pclose(turboPipe);

    return 0;
}

After that is was a matter of chown'ing the binary to belong to root and giving it the setuid bit.  I then adjusted my xmobarrc accordingly:

Config { font = "-*-terminus-*-r-*-*-*-*-*-*-*-*-*-u"
       , bgColor = "black"
       , fgColor = "grey"
       , position = Top
       , lowerOnStart = True
       , commands = [ Run Weather "KCVG" ["-t","<tempF>F","-L","54","-H","80","--normal","green","--high","red","--low","white"] 36000
                , Run Com "uname" ["-r"] "kern" 36000
                , Run Date "%m/%d/%y %H:%M" "date" 100
                , Run Com "sh" ["~/bin/mpd.sh"] "mpd" 25
                , Run Com "~/bin/turboinfo" [""] "cpu" 25
                , Run StdinReader
       ]
       ,sepChar = "%"
       , alignSep = "}{"
              , template = "%StdinReader%}{[<fc=#08FFE2>%mpd%</fc>]|%kern%|%cpu%|%KCVG%|<fc=#ee9a00>%date%</fc>"

Here is the end result:
Xmobar displaying clock frequency post TurboBoost for single threaded loads


Edit: An alternative to this approach may have been to write a shell script to execute on startx which is forked and dumps the output to a fifo generated via mkfifo.  The "info" script executed by xmobar would then pull from the FIFO at an interval that was somewhat close to the update interval passed to turbostat. I prefer the ondemand approach, despite it requiring a lower level of implementation.

5 comments:

  1. You can also view the current frequency in powertop on the Frequency Stats page.

    ReplyDelete
  2. Except that wouldn't satisfy constraint B. But yes powertop does give a lot of c-state info.

    ReplyDelete
  3. XFreq are Widgets to monitor i7 activities, temps, features, bclk ...
    Source code to compile @ http://code.google.com/p/xfreq
    Regards
    CyrIng

    ReplyDelete
    Replies
    1. Enticing, the CLI versions might be a useful addition to my xmobars. Full blown Conky / Gkrellm style system monitors don't really work well for tiling window managers.

      Delete