ntpd: decrease MIN_FREQHOLD by 2, increase "penalty" for largish offset x2

> 2018-07-25:
> ntpd: increase MIN_FREQHOLD by 3
> This means we'll start correcting frequency ~5 minutes after start,
> not ~3.5 ones.
> With previous settings I still often see largish ~0.7s initial offsets
> only about 1/2 corrected before frequency correction kicks in,
> resulting in ~200ppm "correction" which is then slowly undone.

Review of real-world results of the above shows that with small
initial offsets, freq correction can be allowed to kick in sooner,
whereas with large (~0.8s) offsets, we still start freq correction
a bit too soon.

Let's rebalance this a bit.

Signed-off-by: Denys Vlasenko <vda.linux@googlemail.com>
This commit is contained in:
Denys Vlasenko 2019-10-30 12:13:46 +01:00
parent 6b4960155e
commit ea096d6c13

View File

@ -164,7 +164,7 @@
*/
#define INITIAL_SAMPLES 4 /* how many samples do we want for init */
#define MIN_FREQHOLD 12 /* adjust offset, but not freq in this many first adjustments */
#define MIN_FREQHOLD 10 /* adjust offset, but not freq in this many first adjustments */
#define BAD_DELAY_GROWTH 4 /* drop packet if its delay grew by more than this factor */
#define RETRY_INTERVAL 32 /* on send/recv error, retry in N secs (need to be power of 2) */
@ -1876,11 +1876,11 @@ update_local_clock(peer_t *p)
//15:31:53.473 update from:<IP> offset:+0.000007 delay:0.158142 jitter:0.010922 clock drift:+9.343ppm tc:6
//15:32:58.902 update from:<IP> offset:-0.000728 delay:0.158222 jitter:0.009454 clock drift:+9.298ppm tc:6
/*
* This expression would choose MIN_FREQHOLD + 8 in the above example
* (off_065 is +1 for each 0.065536 seconds of offset).
* This expression would choose MIN_FREQHOLD + 14 in the above example
* (off_032 is +1 for each 0.032768 seconds of offset).
*/
unsigned off_065 = abs((int)(tmx.offset >> 16));
G.FREQHOLD_cnt = 1 + MIN_FREQHOLD + off_065;
unsigned off_032 = abs((int)(tmx.offset >> 15));
G.FREQHOLD_cnt = 1 + MIN_FREQHOLD + off_032;
}
G.FREQHOLD_cnt--;
tmx.status |= STA_FREQHOLD;