Message-Id: <1241471016.7444.80.camel@localhost.localdomain>
Date:	Mon, 04 May 2009 14:03:36 -0700
From:	john stultz <johnstul@...ibm.com>
To:	lkml <linux-kernel@...r.kernel.org>
Cc:	Roman Zippel <zippel@...ux-m68k.org>,
	Thomas Gleixner <tglx@...utronix.de>,
	Clark Williams <williams@...hat.com>, linux@...izon.com,
	Ulrich Windl <ulrich.windl@...uni-regensburg.de>
Subject: [RFC][PATCH] Adjust SHIFT_PLL to improve NTP convergence.

With the v2.6.19 Linux kenrel, Roman converted the in-kernel NTP code to
the NTPv4 reference model. 

http://git.kernel.org/?p=linux/kernel/git/torvalds/linux-2.6.git;a=commitdiff;h=f19923937321244e7dc334767eb4b67e0e3d5c74

This was great, because along with his other patches in 2.6.19, the
in-kernel behavior better matched what the userland v4 ntp daemon (used
in most distributions) expects.

However, with this change, we also saw NTP convergence times increase.
This was noticed at the time:
http://lkml.indiana.edu/hypermail/linux/kernel/0609.2/1348.html

But as NTP still functioned (just slower), the explanation of improved
clock stability seemed fair:
http://lkml.indiana.edu/hypermail/linux/kernel/0609.3/0433.html


Since then, I've continued to hear some occasional grumbling on the
topic, but most folks seemed to be getting by ok.

However, more recently, as folks have been upgrading kernels, I've been
involved in a handful of customer issues relating to slow NTP
convergence. It ends up that the current stiffness of the NTP
convergence rate is actually causing problems with NTP clients staying
in close sync when thermal environments change (ie: AC kicking on).

Further, on a fresh system deploy, the current kernel can cause NTPd to
take over 12 hours to find the proper freq value to keep the box in
close sync. Additionally, if one is using the TSC, on reboot the TSC
calibration can often see ~30-60ppm variation, which can cause large
time offsets until NTP converges on the new drift freq.

I started comparing NTP behavior between 2.6.18 and the current kernel,
and the difference in convergence time is very apparent.

In my test, I let the systems sync up, turned NTPd off, set the drift
value to -500ppm, and started NTPd up again. I then ran until NTP
converged and charted the results.

In the first attached image (2.6.30-unmodified.png) you can see 2.6.18's
behavior vs 2.6.30-rc1.

o 2.6.18 very quickly adjusts the drift value and the offset quickly
converges. 

o 2.6.30-rc1 is very slow in adjusting the drift freq, so the offset
grows and grows, until it hits the slew boundary and time is set back to
the time server. This happens a few times until the drift value is close
enough that NTPd can slew the time into sync without forcing the clock
down.

I then modified the SHIFT_PLL value (which defines how stiff NTP's freq
correction is), setting it to 2, and created the second attached image
(2.6.30-pll2.png), comparing against the 2.6.18 numbers in the first
chart.

o Here we see 2.6.30-rc1 matches 2.6.18's behavior very closely.

Note: The freq corrections between kernel versions differ, so the ppm
lines in the chart are expected to be different. However, the rate of
change in both kernels are very very similar.

In reading up about the SHIFT_PLL value, I found the following
discussion:
https://lists.ntp.org/pipermail/hackers/2008-January/003487.html

Where David Mills clarified that the SHIFT_PLL value of 4 in the
nanokernel reference model was appropriate for other Unix systems with
HZ values of 100, and should be reduced as HZ increases.

Linux's in-kernel clock steering is for the most part HZ independent
(calculations are done at the second interval, and then spread out over
the "tick length" which may or may not be connected to HZ). So the rule
of thumb in the link above for setting SHIFT_PLL doesn't really apply.

Thus I ran some experiments and established SHIFT_PLL=2 as having
similar convergence behavior to kernels prior to 2.6.19. In testing with
a few customers, all customers reported much improved NTP convergence
times and also saw improved sync through environmental temperature
variations.

I also ran a fair amount of tests with HZ=100, HZ=1000, NO_HZ, as well
as with changes in various ntp.conf settings, and saw no regressions in
long term clock stability.

As always, feedback or testing (especially on non-x86 arches) would be
greatly appreciated.

thanks
-john


The conversion to the ntpv4 reference model
(f19923937321244e7dc334767eb4b67e0e3d5c74) in 2.6.19 added nanosecond
resolution the adjtimex interface, but also changed the "stiffness" of
the frequency adjustments, causing NTP convergence time to greatly
increase.

SHIFT_PLL, which reduces the stiffness of the freq adjustments, was
designed to be inversely linked to HZ, and the reference value of 4 was
designed for Unix systems using HZ=100. However Linux's clock steering
code mostly independent of HZ. 

So this patch reduces the SHIFT_PLL value from 4 to 2, which causes NTPd
behavior to match kernels prior to 2.6.19, greatly reducing convergence
times, and improving close synchronization through environmental thermal
changes.

The patch also changes some l's to L's in nearby code  to avoid
misreading 50l as 501.

Signed-off-by: John Stultz <johnstul@...ibm.com>

diff --git a/include/linux/timex.h b/include/linux/timex.h
index aa3475f..0daf961 100644
--- a/include/linux/timex.h
+++ b/include/linux/timex.h
@@ -170,17 +170,37 @@ struct timex {
 #include <asm/timex.h>
 
 /*
- * SHIFT_KG and SHIFT_KF establish the damping of the PLL and are chosen
- * for a slightly underdamped convergence characteristic. SHIFT_KH
- * establishes the damping of the FLL and is chosen by wisdom and black
- * art.
+ * SHIFT_PLL is used as a dampening factor to define how much we
+ * adjust the frequency correction for a given offset in PLL mode.
+ * It also used in dampening the offset correction, to define how
+ * much of the current value in time_offset we correct for each
+ * second. Changing this value changes the stiffness of the ntp
+ * adjustment code. A lower value makes it more flexible, reducing
+ * NTP convergence time. A higher value makes it stiffer, increasing
+ * convergence time, but making the clock more stable.
  *
- * MAXTC establishes the maximum time constant of the PLL. With the
- * SHIFT_KG and SHIFT_KF values given and a time constant range from
- * zero to MAXTC, the PLL will converge in 15 minutes to 16 hours,
- * respectively.
+ * In David Mills' nanokenrel reference implmentation SHIFT_PLL is 4.
+ * However this seems to increase convergence time much too long.
+ *
+ * https://lists.ntp.org/pipermail/hackers/2008-January/003487.html
+ *
+ * In the above mailing list discussion, it seems the value of 4
+ * was appropriate for other Unix systems with HZ=100, and that
+ * SHIFT_PLL should be decreased as HZ increases. However, Linux's
+ * clock steering implementation is HZ independent.
+ *
+ * Through experimentation, a SHIFT_PLL value of 2 was found to allow
+ * for fast convergence (very similar to the NTPv3 code used prior to
+ * v2.6.19), with good clock stability.
+ *
+ *
+ * SHIFT_FLL is used as a dampening factor to define how much we
+ * adjust the frequency correction for a given offset in FLL mode.
+ * In David Mills' nanokenrel reference implmentation SHIFT_PLL is 2.
+ *
+ * MAXTC establishes the maximum time constant of the PLL.
  */
-#define SHIFT_PLL	4	/* PLL frequency factor (shift) */
+#define SHIFT_PLL	2	/* PLL frequency factor (shift) */
 #define SHIFT_FLL	2	/* FLL frequency factor (shift) */
 #define MAXTC		10	/* maximum time constant (shift) */
 
@@ -192,10 +212,10 @@ struct timex {
 #define SHIFT_USEC 16		/* frequency offset scale (shift) */
 #define PPM_SCALE ((s64)NSEC_PER_USEC << (NTP_SCALE_SHIFT - SHIFT_USEC))
 #define PPM_SCALE_INV_SHIFT 19
-#define PPM_SCALE_INV ((1ll << (PPM_SCALE_INV_SHIFT + NTP_SCALE_SHIFT)) / \
+#define PPM_SCALE_INV ((1LL << (PPM_SCALE_INV_SHIFT + NTP_SCALE_SHIFT)) / \
 		       PPM_SCALE + 1)
 
-#define MAXPHASE 500000000l	/* max phase error (ns) */
+#define MAXPHASE 500000000L	/* max phase error (ns) */
 #define MAXFREQ 500000		/* max frequency error (ns/s) */
 #define MAXFREQ_SCALED ((s64)MAXFREQ << NTP_SCALE_SHIFT)
 #define MINSEC 256		/* min interval between updates (s) */


Download attachment "2.6.30-unmodified.png" of type "image/png" (42868 bytes)

Download attachment "2.6.30-pll2.png" of type "image/png" (37120 bytes)

Powered by blists - more mailing lists