Fine-grained synchronization on modern networks
in [Embedded]
|
From: Leon on 1 Jan 2010 13:22 On 23 Dec 2009, 17:08, D Yuniskis <not.going.to...(a)seen.com> wrote: > Hi, > > Modern network infrastructure uses *lots* of buffering; > memory is (now) cheap enough to embed throughout the network > fabric. > > With that, fine-grained synchronization over (wired) networks > becomes problematic -- there's no deterministic way for a > processor in a particular node to have any idea of its > relative packet time wrt any other node in the network > (though it is pretty obvious that a packet arrives at > its destination some time *after* leaving its source! :> ) > > Sure, things like NTP *try* to quantify this skew. But, > its goals are much more long-term... if it is wrong on > the short term, there is no significant consequence. > (I also suspect the apparent precision and accuracy that > NTP provides is largely delusional :-/ ) > > So, how *do* you achieve fine-grained synchronization > nowadays? What is *practical*? And theoretically > *achievable* (without an a priori characterization > of the network infrastructure and topology)? XMOS chips and the XLinks that they use for inter-core and inter-chip comms are completely deterministic: http:www.xmos.com Leon
From: Paul Carpenter on 1 Jan 2010 15:20 In article <ec8c6001-f408-43d9-97fc- 0cacf4201dfc(a)d20g2000yqh.googlegroups.com>, leon355(a)btinternet.com says... > On 23 Dec 2009, 17:08, D Yuniskis <not.going.to...(a)seen.com> wrote: > > Hi, > > > > Modern network infrastructure uses *lots* of buffering; > > memory is (now) cheap enough to embed throughout the network > > fabric. > > > > With that, fine-grained synchronization over (wired) networks > > becomes problematic -- there's no deterministic way for a > > processor in a particular node to have any idea of its > > relative packet time wrt any other node in the network > > (though it is pretty obvious that a packet arrives at > > its destination some time *after* leaving its source! :> ) > > > > Sure, things like NTP *try* to quantify this skew. But, > > its goals are much more long-term... if it is wrong on > > the short term, there is no significant consequence. > > (I also suspect the apparent precision and accuracy that > > NTP provides is largely delusional :-/ ) > > > > So, how *do* you achieve fine-grained synchronization > > nowadays? What is *practical*? And theoretically > > *achievable* (without an a priori characterization > > of the network infrastructure and topology)? > > XMOS chips and the XLinks that they use for inter-core and inter-chip > comms are completely deterministic: Which has nothing to do with "embed throughout the network fabric" unless the network fabric is VERY small. Obviously, you use full TCP/IP between these chips so that you can run "things like NTP *try* to quantify this skew" -- Paul Carpenter | paul(a)pcserviceselectronics.co.uk <http://www.pcserviceselectronics.co.uk/> PC Services <http://www.pcserviceselectronics.co.uk/fonts/> Timing Diagram Font <http://www.gnuh8.org.uk/> GNU H8 - compiler & Renesas H8/H8S/H8 Tiny <http://www.badweb.org.uk/> For those web sites you hate
From: Leon on 1 Jan 2010 17:44 On 1 Jan, 20:20, Paul Carpenter <p...(a)pcserviceselectronics.co.uk> wrote: > In article <ec8c6001-f408-43d9-97fc- > 0cacf4201...(a)d20g2000yqh.googlegroups.com>, leon...(a)btinternet.com > says... > > > > > On 23 Dec 2009, 17:08, D Yuniskis <not.going.to...(a)seen.com> wrote: > > > Hi, > > > > Modern network infrastructure uses *lots* of buffering; > > > memory is (now) cheap enough to embed throughout the network > > > fabric. > > > > With that, fine-grained synchronization over (wired) networks > > > becomes problematic -- there's no deterministic way for a > > > processor in a particular node to have any idea of its > > > relative packet time wrt any other node in the network > > > (though it is pretty obvious that a packet arrives at > > > its destination some time *after* leaving its source! :> ) > > > > Sure, things like NTP *try* to quantify this skew. But, > > > its goals are much more long-term... if it is wrong on > > > the short term, there is no significant consequence. > > > (I also suspect the apparent precision and accuracy that > > > NTP provides is largely delusional :-/ ) > > > > So, how *do* you achieve fine-grained synchronization > > > nowadays? What is *practical*? And theoretically > > > *achievable* (without an a priori characterization > > > of the network infrastructure and topology)? > > > XMOS chips and the XLinks that they use for inter-core and inter-chip > > comms are completely deterministic: > > Which has nothing to do with > > "embed throughout the network fabric" > > unless the network fabric is VERY small. > > Obviously, you use full TCP/IP between these chips so that you can run > > "things like NTP *try* to quantify this skew" They have Ethernet connectivity as well, with TCP/IP stacks available. They are fast enough to implement Ethernet in software. Leon
From: Paul Carpenter on 1 Jan 2010 19:13 In article <1d581e31-5c65-4add-ad9e- 8ec7d599b3c1(a)o28g2000yqh.googlegroups.com>, leon355(a)btinternet.com says... > On 1 Jan, 20:20, Paul Carpenter <p...(a)pcserviceselectronics.co.uk> > wrote: > > In article <ec8c6001-f408-43d9-97fc- > > 0cacf4201...(a)d20g2000yqh.googlegroups.com>, leon...(a)btinternet.com > > says... > > > On 23 Dec 2009, 17:08, D Yuniskis <not.going.to...(a)seen.com> wrote: > > > > Hi, > > > > Modern network infrastructure uses *lots* of buffering; > > > > memory is (now) cheap enough to embed throughout the network > > > > fabric. > > > > > > With that, fine-grained synchronization over (wired) networks > > > > becomes problematic -- there's no deterministic way for a > > > > processor in a particular node to have any idea of its > > > > relative packet time wrt any other node in the network > > > > (though it is pretty obvious that a packet arrives at > > > > its destination some time *after* leaving its source! :> ) > > > > > > Sure, things like NTP *try* to quantify this skew. But, > > > > its goals are much more long-term... if it is wrong on > > > > the short term, there is no significant consequence. > > > > (I also suspect the apparent precision and accuracy that > > > > NTP provides is largely delusional :-/ ) > > > > > > So, how *do* you achieve fine-grained synchronization > > > > nowadays? What is *practical*? And theoretically > > > > *achievable* (without an a priori characterization > > > > of the network infrastructure and topology)? > > > > > XMOS chips and the XLinks that they use for inter-core and inter-chip > > > comms are completely deterministic: > > > > Which has nothing to do with > > > > "embed throughout the network fabric" > > > > unless the network fabric is VERY small. > > > > Obviously, you use full TCP/IP between these chips so that you can run > > > > "things like NTP *try* to quantify this skew" > > They have Ethernet connectivity as well, with TCP/IP stacks available. > They are fast enough to implement Ethernet in software. Whoosh... Leon <---------------------------------------> point -- Paul Carpenter | paul(a)pcserviceselectronics.co.uk <http://www.pcserviceselectronics.co.uk/> PC Services <http://www.pcserviceselectronics.co.uk/fonts/> Timing Diagram Font <http://www.gnuh8.org.uk/> GNU H8 - compiler & Renesas H8/H8S/H8 Tiny <http://www.badweb.org.uk/> For those web sites you hate
First
|
Prev
|
Pages: 1 2 3 4 Prev: OpenTherm protocol: looking for schematics Next: From Access to MySQL |