[j-nsp] PTP design
wojciech at owczarek.co.uk
Wed Sep 21 08:15:22 EDT 2016
Firstly, there is no 1 microsecond requirement. MiFID II states 100
microseconds, and only for low-latency environments.
- Best design: any redundant network design. Specifically for PTP: make
sure you have more than one PTP grandmaster (GM). This is a long topic.
- Your best bet is to use multicast since this is a Finance environment.
Some equipment in this space only supports multicast PTP.
- Yes, there are scalability issues, PTP grandmasters have limited message
throughput. This comes down to x slaves (clients) maximum (assuming some
n/sec message rate). They are not routers. Speak to your (prospective)
vendor about the limitations.
- Only one multicast group, 22.214.171.124. This is an IEEE standard, not
IETF. I really suggest to get a copy of the standard. You will need it. PTP
uses two UDP ports: 319 (for all timestamped messages) and 320
(non-timestamped). PTP communicates in two directions - this is especially
important for multicast. GM sends Sync and Announce messages to slaves,
slaves send Delay Request to which GM replies with Delay Response.
- You don't have to dedicate a VLAN, but you can, and it's a good idea. As
to delivering PTP to the host, some NICs and PTP-aware switches don't
readily deal with VLAN tags for PTP, so even if your hosts use tagged
traffic, the PTP connection should probably be untagged.
- It all depends on your network load, but unless your switches are PTP
aware (like PTP Transparent Clock), you don't want to run PTP over a busy
network as it impacts packet delay variation. This takes you back to the
design - if network is not PTP aware you may want to put a parallel design
in place that only delivers timing.
To put it simply, deploying PTP is nowhere as easy as deploying NTP.
Luckily the MiFID requirement isn't that stringent, but it's best to aim
for better than 100 us, because chances are it will go lower before your
next h/w refresh (unless you refresh very often). Get learning, now.
There are some time sync related conferences every year, the next European
one is ITSF which is in Prague this year in November. There will be a
session devoted to finance / mifid timing.
http://itsf2016.executiveindustryevents.com/Event/home if you're
interested. There is also a tutorial day. Other events are WSTS and ISPCS.
On 20 September 2016 at 18:01, james list <jameslist72 at gmail.com> wrote:
> Hi experts!
> More than a vendor related question I’m wondering to discuss or get hints
> regarding the upcoming mifid2 new PTP request (max divergence from UTC of 1
> microsecond) implementation that will be requested since Jan 2018.
> I’d like to setup in my DC two fully redundant PTP source, for this reason
> I’m planning to use two different antennas, coax and supplier, but here the
> first doubts:
> - What is best design to provide redundancy ?
> - Do I have to use multicast or unicast (like NTP) ?
> - Is there any scalability issue ?
> - If I use multicast, which are the multicast group used by the
> PTP vendors ? Is there any ietf assigned group ?
> - Do I have to dedicate a single dedicated vlan where the server
> has to connect to get the multicast packets ?
> I’ve many doubts and maybe we can share some commons ideas if anybody else
> is going to setup the service…
> Thanks in advance for any feedback
> juniper-nsp mailing list juniper-nsp at puck.nether.net
More information about the juniper-nsp