edit - for Windows 10 or Server 2016 devices there is now a definitive, native solution available - see this answer
Depending on HOW time-critical your applications are, one option could be to configure NT5DS rather than NTP, if these two domain member servers are not currently using NT5DS.
Considering your issues are on the order of seconds, this may not help. But before you head out to install another solution, it may be worth investigating. If it's still too slow, see ErikE's answer.
We were having issues with our Domain Controllers at two sites being nearly 3 minutes out of sync. Kerberos was still functional, but some ancillary applications using AD authentication were not. The issue was that two DC's were set to NTP (pulling from the SAME time source) and the rest were set to NT5DS. Setting one DC to NT5DS (leaving one authoritative domain time source) resolved this issue.
On restarting their time services, all remaining DC's and member servers fell in line to the minute (didn't check to the seconds, since we didn't need it) with the single DC now acting as the domain time source (& pulling via NTP from a higher stratum source)
This allows the (tenuous) inference that NT5DS may do a better job keeping times between servers in the same domain consistent than NTP does.
To configure a domain member server with NT5DS:
w32tm /config /syncfromflags:domhier /update
Then restart the time service:
net stop w32time && net start w32time