I have my own cluster (intranet! no outside connection) which i use it for distributed programming purposes. I need to synchronise time of the cluster , therefore i run ntp server on one of my server. I need the time to be accurate on the order of micro seconds however i found out that it is not that accurate! i.e one machine is behind other...

Communication medium is gigabit ethernet and pcs are 2.60Ghz each...

I found out that there is a 100-10000 micro seconds gap among them ... is it possible to tune it better ?

broadcastdelay is set to 0.0005
cheers

Time will probably never be entirely synchronized, but when you say micro seconds, how many are you thinking? Perhaps you could post a general configuration file, a diagram of the ntp setup (what are the ntp strata). What is your polling time? Are you using multicasting? Have you collected stats? Where does the main ntp server get its time? etc. etc.

1 members found this post helpful.

well in the order of 10 micro seconds would do what i need to do. I am trying to synchronise the workstations to my own server's time. So there is one server and workstations connected to it via gibabit switch. This network is for simulation purposes...

ntp.conf:

client:
server 10.0.0.100
broadcastclient yes
broadcastdelay 0.0005
keys /etc/ntp.keys

server:
server 127.127.1.0
fudge 127.127.1.0 stratum 10
broadcast 10.0.0.255
keys /etc/ntp.keys

"...in the order of 10 micro seconds would do..."

Yeeowtch! You don't ask for much, do you?

I'm not at all sure that the clocks in most PCs can be trusted to a microsecond, which will make keeping them synchronized to a tolerance of around 10 microseconds pretty tough.

The NTP protocol is designed to keep time between servers spread out geographically, and it's time as humans see it -- where a second is a fairly short period of time. The documentation says it keeps accuracy to within a few "tens of milliseconds" on LANs.

If you really and truly need synchronization that tight, you might ought to start looking for special-purpose hardware and software.

All that being said, there is one thing you can do to improve things.

Lets say you have five client machines, 10.0.0.101 through 105. On each client, your server list should look like:

server 10.0.0.100
server 10.0.0.101
server 10.0.0.102
server 10.0.0.103
server 10.0.0.104
server 10.0.0.105

Because of the way NTP works, the more machines it can talk to (that are synchronized themselves) the more accurate it is. (Most of the fanciness of the logic is in trying to compensate for transmission delays, and the more servers it has to compare to each other, the more accurate it's guesses are.)

Leave the server config as you have it. It's going to be the source of time for the cluster.

Since you'll have a peak of network activity when ntp starts up, you might be better off to have each client machine point to three or four other clients (in addition to the server), but not to every single client. I really don't know which will work best -- you'll just have to experiment.

One thing about NTP that may cause you some problems:

It normally starts out checking once every 64 seconds, and each time the next check is "close enough", it doubles the interval, maxing out at 1024 seconds. Problem is, for your purposes, its definition of "close enough" probably ISN'T going to be close enough.

You might be better off running ntpdate. Another thing you could do with ntpdate is put a random few seconds delay in the script on each machine, so that you avoided all the congestion from everybody trying to sync time at once.

Whoops, just thought of something. If you're not running the NTP daemon, the client can't act a server, either to another NTP daemon or ntpdate. So the ntpdate option is out.

If adding more servers doesn't help, you're either looking at doing some pretty heavy-duty hacking on NTP itself, or you're going to have to find another product.

Good luck,

CHL

1 members found this post helpful.

I know this post is old enough for the original poster to find a solution but for anyone like myself I'll add that there is a possibilty to make your own stratum 1 server as has been suggested here:
http://www.ntp.org/ntpfaq/NTP-s-algo.htm
With cheap gps dongle and gpsd this should'n be hard/expensive to set up.
I haven't done this myself, so I can't provide any more info, but this might be a way to go.
https://duckduckgo.com/?q=build+stra...ps&t=canonical

I wish you success!
Aas.