[PATCH] Handled no difference in seek times

[PATCH] Handled no difference in seek times

[Alan D. Brunelle]

For some reason recent kernels (2.6.25.4, for example) have lost a lot
of resolution in our blktrace times. This can result in lots of things
happening "simultaneously." This change at least tries to handle the
case where all the seeks happen at once.

Probably have other issues that need to be looked into...

[[Jens: I've pushed this, so you don't need to apply this, but I'm
wondering have others seen this? I'm seeing 0.01 - which I think means
we're gated by HZ (set to 100 on this .config, I'm trying w/ HZ\x1000 to
check this theory). A /lot/ happens in 1/100th of a second on modern
computers! :-) ]]

Signed-off-by: Alan D. Brunelle <alan.brunelle@hp.com>
---
 btt/doc/btt.tex |    4 +++-
 btt/seek.c      |    7 ++++---
 2 files changed, 7 insertions(+), 4 deletions(-)

diff --git a/btt/doc/btt.tex b/btt/doc/btt.tex
index 9938f67..87e7082 100644
--- a/btt/doc/btt.tex
+++ b/btt/doc/btt.tex
@@ -769,7 +769,9 @@ Device:       rrqm/s   wrqm/s     r/s     w/s
rsec/s    wsec/s

   When there is only a single data point within a 1-second window,
   \texttt{btt} will just output the time value for the point, and the
-  value 1.0 in the second column.
+  value 1.0 in the second column. If there is no perceived difference
+  in the times present for the current sample, then the second columns
+  value is the number of seeks present at that time.

   Otherwise, if $\alpha$ and $\Omega$ are the first and last times
   seen within a 1-second window, and $\nu$ are the number of seeks seen
diff --git a/btt/seek.c b/btt/seek.c
index fec57c4..69400c8 100644
--- a/btt/seek.c
+++ b/btt/seek.c
@@ -18,6 +18,7 @@
  *  Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA
02111-1307  USA
  *
  */
+#include <float.h>
 #include "globals.h"

 static struct file_info *seek_files = NULL;
@@ -102,13 +103,13 @@ static void sps_emit(struct seeki *sip)
 {
 	double tstamp, s_p_s;
 	struct sps_bkt *sps = &sip->sps;
+	double delta = sps->t_last - sps->t_start;

-	if (sps->nseeks = 1) {
-		s_p_s = 1.0;
+	if ((sps->nseeks = 1) || (delta < DBL_EPSILON)) {
+		s_p_s = (double)(sps->nseeks);
 		tstamp = sps->t_start;
 	}
 	else {
-		double delta = sps->t_last - sps->t_start;

 		s_p_s = (double)(sps->nseeks) / delta;
 		tstamp = sps->t_start + (delta / 2);
-- 
1.5.4.3

Re: [PATCH] Handled no difference in seek times

[M. Edward (Ed) Borasky]

Alan D. Brunelle wrote:
> For some reason recent kernels (2.6.25.4, for example) have lost a lot
> of resolution in our blktrace times. This can result in lots of things
> happening "simultaneously." This change at least tries to handle the
> case where all the seeks happen at once.
> 
> Probably have other issues that need to be looked into...
> 
> [[Jens: I've pushed this, so you don't need to apply this, but I'm
> wondering have others seen this? I'm seeing 0.01 - which I think means
> we're gated by HZ (set to 100 on this .config, I'm trying w/ HZ\x1000 to
> check this theory). A /lot/ happens in 1/100th of a second on modern
> computers! :-) ]]
> 
> Signed-off-by: Alan D. Brunelle <alan.brunelle@hp.com>
> ---
>  btt/doc/btt.tex |    4 +++-
>  btt/seek.c      |    7 ++++---
>  2 files changed, 7 insertions(+), 4 deletions(-)
> 
> diff --git a/btt/doc/btt.tex b/btt/doc/btt.tex
> index 9938f67..87e7082 100644
> --- a/btt/doc/btt.tex
> +++ b/btt/doc/btt.tex
> @@ -769,7 +769,9 @@ Device:       rrqm/s   wrqm/s     r/s     w/s
> rsec/s    wsec/s
> 
>    When there is only a single data point within a 1-second window,
>    \texttt{btt} will just output the time value for the point, and the
> -  value 1.0 in the second column.
> +  value 1.0 in the second column. If there is no perceived difference
> +  in the times present for the current sample, then the second columns
> +  value is the number of seeks present at that time.
> 
>    Otherwise, if $\alpha$ and $\Omega$ are the first and last times
>    seen within a 1-second window, and $\nu$ are the number of seeks seen
> diff --git a/btt/seek.c b/btt/seek.c
> index fec57c4..69400c8 100644
> --- a/btt/seek.c
> +++ b/btt/seek.c
> @@ -18,6 +18,7 @@
>   *  Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA
> 02111-1307  USA
>   *
>   */
> +#include <float.h>
>  #include "globals.h"
> 
>  static struct file_info *seek_files = NULL;
> @@ -102,13 +103,13 @@ static void sps_emit(struct seeki *sip)
>  {
>  	double tstamp, s_p_s;
>  	struct sps_bkt *sps = &sip->sps;
> +	double delta = sps->t_last - sps->t_start;
> 
> -	if (sps->nseeks = 1) {
> -		s_p_s = 1.0;
> +	if ((sps->nseeks = 1) || (delta < DBL_EPSILON)) {
> +		s_p_s = (double)(sps->nseeks);
>  		tstamp = sps->t_start;
>  	}
>  	else {
> -		double delta = sps->t_last - sps->t_start;
> 
>  		s_p_s = (double)(sps->nseeks) / delta;
>  		tstamp = sps->t_start + (delta / 2);

I just upgraded to 2.6.25.4 -- is there something I should try?

I am running Gentoo Linux and collecting data on various things like 
kernel builds, recompiling PHP, running a Windows guest with VMware 
Workstation 6, and if I can get it working in the next day or so, a Ruby 
on Rails benchmark.

So far everything I've tried has turned out to be processor bound 
(Athlon64 X2 2.2 GHz dual-core with 4 GB of RAM), so I may boot with 
only 1 GB to try to get the disk utilization up. :)

Re: [PATCH] Handled no difference in seek times

[Alan D. Brunelle]

Hm, might just be my testbed - I'm getting:

[    6.665781] Clocksource tsc unstable (delta = 88002879 ns)

in dmesg, and looking at the code I /think/ this drops the high res
timer off.

<sigh>
Alan