Resolved The problem was Hyper-V on that machine. I removed Hyper-V, installed VMware Server, ran the same VM. Time sync issues went away (< 100ms difference after a day).

My setup is like this:

HYV1 - HyperV machine (non domain) - sync irrelevant
AD1  - VM AD server on HYV1, sync'd to time.nist.gov. HyperV time sync off.
S1   - Physical machine, sync'd to domain. 
S2   - Physical machine running HyperV, sync'd to domain.
V1   - Linux VM machine on S2, sync'd to AD1. No HyperV integration.

AD1 and S1 have fine sync -- stripchart shows less than 100ms difference.

S2 drifts like crazy. Here's a bit of the stripchart against AD1:

18:33:22 d:+00.0010138s o:+05.4101899s 
18:33:24 d:+00.0010138s o:+05.4319765s 
18:33:26 d:+00.0000000s o:+05.4788429s 
18:33:28 d:+00.0000000s o:+05.6089942s 
18:33:30 d:+00.0010138s o:+05.7240269s 
18:33:32 d:+00.0000000s o:+06.0421911s 
18:33:34 d:+00.0081104s o:+06.5613708s 
18:33:37 d:+00.0000000s o:+06.9096594s 
18:33:39 d:+00.0000000s o:+06.8867838s 
18:33:41 d:+00.0010127s o:+06.8936401s

In 20 seconds, it drifted over a second. If I manually reset it to within 1s, within a few minutes it'll be back drifting about 2 seconds. Overnight it went from ~2s to ~5s. The Linux VM inside S2 has perfect sync with AD1.

Here's the config:

C:\Users\mgg>w32tm /dumpreg /subkey:Parameters

Value Name                 Value Type          Value Data
------------------------------------------------------------

ServiceDll                 REG_EXPAND_SZ       %systemroot%\system32\w32time.dll
ServiceMain                REG_SZ              SvchostEntry_W32Time
ServiceDllUnloadOnStop     REG_DWORD           1
Type                       REG_SZ              NT5DS
NtpServer                  REG_SZ              ad01.mydomain ad02.mydomain


C:\Users\mgg>w32tm /dumpreg /subkey:Config

Value Name                Value Type          Value Data
-----------------------------------------------------------

FrequencyCorrectRate      REG_DWORD           4
PollAdjustFactor          REG_DWORD           5
LargePhaseOffset          REG_DWORD           50000000
SpikeWatchPeriod          REG_DWORD           900
LocalClockDispersion      REG_DWORD           9
HoldPeriod                REG_DWORD           5
PhaseCorrectRate          REG_DWORD           1
UpdateInterval            REG_DWORD           30000
EventLogFlags             REG_DWORD           2
AnnounceFlags             REG_DWORD           5
TimeJumpAuditOffset       REG_DWORD           28800
MinPollInterval           REG_DWORD           2
MaxPollInterval           REG_DWORD           8
MaxNegPhaseCorrection     REG_DWORD           -1
MaxPosPhaseCorrection     REG_DWORD           -1
MaxAllowedPhaseOffset     REG_DWORD           300

I looked at the event log, and apart from warnings about sync (after it gets way out of sync), there's no other warnings.

How can I go about troubleshooting this? It's the only machine that is having this problem. All the other machines (physical and virtual) are doing fine.

Edit: To clarify: The VM (AD1) has integration turned off and syncs to time.nist.gov. AD1 is fine. It's the physical machine S1 that can't sync to AD1 and drifts all over. All the other physical servers are able to sync to AD1 just fine.

Update So, it appears to be an issue of running the VM. The clock slips slowly with the VM off. Turned on, it immediately starts losing seconds. I swt the VM to only use half the resources, and that seems to have slightly mitigated it, for now. Thanks!