Delivering Better Time-of-Day Using Synchronous Ethernet and Yaakov (J) Stein, Alon Geva, Gabriel Zigelboim RAD Data Communications

Transcription

1 Delivering Beer Time-of-Day Using Synchronous Eherne and 1588 Yaakov (J) Sein, Alon Geva, Gabriel Zigelboim RAD Daa Communicaions

2 The problem I wan discuss he use of 1588 (or NTP for ha maer) for ime of day (wall-clock) disribuion in conjuncion wih Synchronous Eherne (Sync-E) Sync-E is a physical layer frequency disribuion mechanism Wha does Sync-E have o do wih for ime disribuion? In pracice, frequency and ime disribuion are ofen inexricably inerwined So, ime disribuion proocols (1588 or NTP) always do frequency disribuion as well SyncE_1588 Slide

3 When don' we need a frequency disribuion proocol for ime-of-day? If we have a high accuracy local frequency source high ime updae rae Non-sringen ime accuracy requiremen Then we don' need o disribue frequency o obain ime If, for example, local oscillaor is an OCXO wih a lifeime accuracy of ppb ime updae rae is 1 per second ime accuracy requiremen is 5 nanoseconds Then i's OK no o updae of he local frequency reference since he ime error due o he frequency offse is always less han nanoseconds SyncE_1588 Slide 3

4 When do we need a frequency disribuion proocol for ime-of-day? The problem arises when we don' have accurae local frequency We can increase he ime disribuion updae rae bu ha increases he (nework, compuaional) resource drain If he local reference is sable (bu no accuraely calibraed) hen we can use a frequency disribuion proocol o se i For example, an inexpensive crysal's frequency accuracy migh be 1 ppm bu afer frequency lock i is wihin 1 ppb and hus is drif is only 1 nanosecond per second So frequency disribuion grealy reduces he resource drain SyncE_1588 Slide 4

5 Frequency disribuion proocols vs. Sync-E Bu using a higher layer ime disribuion proocol for locking frequency has is drawbacks In paricular, i increases convergence ime (ime for error < required) increases seady sae ime error Sync-E is a physical layer frequency disribuion mechanism Sync-E requires special hardware bu pus no furher demands on (BW or compuaion) resources If we have Sync-E already insalled for applicaions ha require phase lock especially insalled since less expensive han beer oscillaor hen we don' need o use higher layer frequency disribuion SyncE_1588 Slide 5

6 Advanages and disadvanage By augmening 1588 wih Sync-E, we can obain: beer seady-sae performance faser convergence o he desired ime accuracy capabiliy of on-demand ime ransfer possibiliy of using lower updae raes during peak hours Bu, if Sync-E is no already insalled: may require a forklif upgrade of all swiches along he pah may hus be prohibiively expensive SyncE_1588 Slide 6

7 Time disribuion over packe swiched nework ime updae packe received (local clock) PDV wander ime disribuion packes are sen from source wih imesamps measured by source clock hese packes are received afer nework propagaion delay ime updae packe sen (source clock) Wihou queuing delay, all packes received afer minimal delay oherwise here is Packe Delay Variaion (PDV) Bu he receiver measures he arrival ime using is local clock and his local clock has wander wih respec o source clock SyncE_1588 Slide 7

8 Two conribuions So he observed arrival imes differ from he imesamps due o wo effecs: he difference in frequency beween local and source clocks he packe propagaion delay, which in urn is made up of he minimal (elecrical) propagaion delay he queuing delay (ime he packe wais because oher raffic in queue) The job of a ime disribuion sysem is o lock he local clock frequency ono he source clock measure he propagaion delay (ranging) Wih Sync-E he firs ask is performed by he physical layer and so he firs effec is minimized SyncE_1588 Slide 8

9 Convergence ime Assume ha our implemenaion firs sabilizes local frequency reference and only aferwards locks he ime Then he ime o converge o desired ime accuracy is he sum of he ime o obain frequency lock he ime for ranging procedure o converge error If we eliminae he firs erm by using Sync-E we cerainly reduce he convergence ime Even if our implemenaion simulaneously adaps frequency and ime is convergence ime is longer since early ime adapaions are relaively meaningless SyncE_1588 Slide 9

10 Seady sae error Once converged, sandard ime disribuion sysems coninue updaing he local frequency reference coninue ranging o updae local ime The ime error is he sum of he frequency error (wander) conribuion he ranging error (caused by PDV, asymmery, ec.) If we eliminae he firs erm by using Sync-E we cerainly reduce he seady-sae error error Sync-E sill leaves some residual frequency error bu physical layer frequency locking is more reliable more accurae han higher layer frequency disribuion SyncE_1588 Slide 1

11 A subler (bu imporan) effec There is ye anoher way Sync-E can help 1588 or NTP Ranging echniques funcion by reques and response exchange compuing round-rip ime using four imesamps esimaing one-way ime assuming symmery minimize PDV effec by minimum gaing (if possible) Minimum gaing assumes ha here are some packes ha raverse nework wih essenially no queuing delay idenifies hese packes by finding minimum round-rip delay If he probabiliy of minimal delay one-way raversal is p hen he probabiliy of round-rip minimal delay raversal is p This probabiliy may be vanishingly small! Bu we don' really need o have a single ransacion wih minimal raversal ime in boh direcions if P=1% updae 1pps hen 1 sec unil one way minima bu 1 sec unil round-rip minimum SyncE_1588 Slide 11

12 Two separae minima We can monior he four imesamps and noe he minimal difference independenly in each direcion he probabiliy of hese wo evens is p no p Were he slave's clock o be locked o he maser clock he regular calculaion could be used Bu when he slave's clock drifs beween he wo minimal raversal evens he calculaion acquires a corresponding error erm Δf By using Sync-E we ensure ha he slave clock is locked hus enabling use of wo separae minimal raversal evens and hus immensely improving he ranging performance ( ) d SyncE_1588 Slide 1

13 Using separae minimum mehod wih IEEE 1588 If slave clock has offse T hen slave ime corresponds o maser ime T If slave also has frequency offse Δf() hen slave ime corresponds o maser ime T - Δf() d OC/BC Maser ime PSN Slave ime OC/BC Timesamps known by slave wihou Sync-E: ( k) ( k) = d ( k) T Δf ( ) 1 1 d -ms 1 Sync Follow_Up d 1 ( k) 1, 4 4 ( k) ( ) = ( ) k d34 k T Δf ( ) d -sm 4 Delay_Req d 34 ( k) 3 1,, 3 ( ) = wih Sync-E we can ake Δf : 4 ( k) 1( k) = d1( k) T ( k) 3( k) = d34( k) + T Delay_Resp 1,, 3, 4 SyncE_1588 Slide 13

14 Theoreic discussion assumpions To be specific, we will assume: Time updae rae 1 PPS in he forward direcion (for Adapive Clock Recovery - ACR) 1 PPS in he backward direcion (for TOD) Adapive clock recovery based solely on forward direcion Adapive clock recovery bandwidh of 1 mhz Adapive clock recovery is based on an OCXO local reference For whie Gaussian nework delay mean delay is assumed o be symmeric For runcaed disribuions minimum delay is assumed o be symmeric SyncE_1588 Slide 14

15 Addiive Gaussian Packe Delay When he PDV disribuion is Gaussian here is no an appreciable proporion of packes wih minimal delay so we rely on he enire se of delay values (no minimum gaing) Assuming symmery, we average over k: 1 [ ( k) ( k) ( ( k) ( k) )] = = 1 d 34 ( k ) ( k) d ( k) + T + Δf () 1 = assuming mean symmery 1 d + ( k ) 4 Δf () Error d Noe: we need o average over a long ime since he firs erm is zero only on average when Δf he las expression conribues unavoidable error o he TOD esimaion Thus, having zero frequency error (Sync-E) makes a difference SyncE_1588 Slide 15

16 Truncaed Packe Delay 5 Cascaded Swiches GE 1 GE 5... GE 8% Background raffic % of packes in each direcion undergo minimum delay ha is approximaely Gaussian wih 1.75 μsec mean and.6 μsec sd-dev (oal delay has 8 μsec sd-dev) σ σ o = min SyncE_1588 Slide 16

17 = Independenly exploiing he min on boh direcions Here we have a porion of he packes experiencing min delay relying on he approximaely % of enire delay values se on each direcion Therefore, any ime offse due o frequency error canno be accuraely measured anymore: min min 1 1 ( k ) m { ( k ) ( k )} = min{ d ( k )} T Δf ( ) m ( k ) 4 n { ( k ) ( k )} = min{ d ( k )} + T + Δf ( ) 4 min n { ( k ) ( k )} min{ ( k ) ( k )} min 4 n ( k ) m { d ( k )} min{ d ( k )} + T + Δf () n m n 3 n m m n m 1 m 1 = d d d + 4 ( k ) = assuming min symmery Error n Δf Again zero frequency error makes a difference! Using Sync-E () d SyncE_1588 Slide 17

18 Seady sae pk-o-pk TOD error performance (for T sec inegraion ime) For he Gaussian case (using ACR) TOD can be approximaed by: TOD pkpk_err σ delay 7 + T ranging error 1mHz 5Hz σ ACR error delay 7 o conver sd o pkpk 1 pps so # of TOD ransacions = 1T For he runcaed case (using ACR) TOD is only approximaely: TOD pkpk_err σ delay T 1mHz 5Hz σ delay inpu-oupu BW raio only 4T since one fifh have min delay ACR s phase error conribuion assuming linear PLL model and consan 8% load Now, applying Sync-E SyncE_1588 Slide 18

19 Asympoic behavior lower bound due o ACR non-perfec frequency lock wih load changes, he effec occurs dramaically earlier! T > order of magniude improvemen abou an order of magniude improvemen SyncE_1588 Slide 19

20 Simulaions We performed a range of simulaions o es our proposal For each simulaion we ran wo ime recovery algorihms full 1588 algorihm - frequency and ime recovery ime-only 1588 algorihm - frequency aken from SyncE The updae raes were aken o be: Maser->Slave - 1pps Slave->Maser - 1pps We ran wo nework scenarios G.861 Tes scenario for wo-way proocols Gaussian PDV SyncE_1588 Slide

21 Simulaion seup (based upon G.861 wo-way esing seup) Reference Timing Signal (PRC/UTC) Packe Delay Variaion Jier, Wander, Frequency and Time accuracy % consan load Reverse channel Traffic Generaor Tes Equipmen Tes Equipmen PEC Server GE Reference poin 1 1 GE GE 3... GE... GE 5 1 FE or GE Reference poin PEC Clien (DUT) Reference poin 3 Clock Disurbance load according o raffic models - forward direcion Disurbance load according o raffic models - reverse direcion Flow of ineres - forward direcion Flow of ineres - reverse direcion Eherne Swiches GE = 1 Gbps Eherne FE = 1 Mbps Eherne Forward channel Traffic Generaor 8% consan load for he Sync-E case each swich conforms o G.86 SyncE_1588 Slide 1

22 G.861 symmeric scenario 14 1 Absolue Time Error G.861 TesCase 1 Model Adapive SyncE Time error (nsec) Time Error [nsec] don' worry abou 4 nsec offse Time [Sec] SyncE_1588 Slide

23 Gaussian PDV scenario 6 4 Absolue Time Error Gaussian Noise, d=4ms σ MS =5μsec σ SM =1μsec Adapive SyncE Time error Error [nsec] (nsec) Time [Sec] SyncE_1588 Slide 3