[ntp:questions] Proposed NTP solution for a network

Ryan Malayter malayter at gmail.com
Tue Mar 3 18:01:38 UTC 2009


On Tue, Mar 3, 2009 at 7:13 AM, Jason <bmwjason at bmwlt.com> wrote:
> The critical time-stamp of the transactions must be tighter than 100us.

As others have stated, this is likely not possible using Ethernet,
especially on blade servers, which often share IO channels. You will
likely see switch queuing latencies higher than this if your LAN is at
all busy or has bursts of activity.

What is the application requirement here? I would also question if you
can even achieve time stamping of 100 *microseconds* repeatably at
your application layer. I am guessing we are talking abuot some sort
of trading application... I would think you would see transaction
init/commit times with much higher jitter than 100 microseconds,
especially if they must be written to disk or across the LAN.

Are you interested in time stamping to ensure absolute ordering of
transactions? If that is the case, typically a 2-phase commit protocol
with an application-layer counter is used, not "system clock time",
which is known to be unreliable. Almost all middleware and database
stacks have primitives for this sort of thing to ensure that
transactions can be ordered properly even in distributed/replicated
environments.

--
RPM



More information about the questions mailing list