QoS in Frame relay networks

Transcription

1 Frame Relay ndrea ianco Telecommunication Network Group Network management and QoS provisioning - 1 Frame Relay: characteristics Packet switching with virtual circuit service Label named LI: ata Link onnection Identifier Virtual circuits are bi-directional onnection is associated with the virtual circuit No error control (L-control is not used even at edge) No flow control LP-F protocol Packet size: variable up to 4096byte Mainly thought for data traffic Network management and QoS provisioning - 2 LPF packet RESS field contains the LI (ata Link onnection Identifier) and some additional bits L-ORE L-ONTROL FLG R (MS) R (LS) control Information R (MS) R (LS) FLG Network management and QoS provisioning - 3 Pag. 1

2 LI: ata Link onnection Identifier FEN/EN: forward/backward explicit congestion notification E: discard eligibility /R: command/response /: LI or L- ORE : extension bit RESS field lower LI upper LI /R FEN EN E upper LI /R LI FEN EN E lower LI or L-ORE control / upper LI /R 0 LI FEN EN E 0 LI 0 lower LI or L-ORE control / 1 efault format (2 byte) 3 byte format 4 byte format Network management and QoS provisioning - 4 Frame Relay: user-network interface Negotiable parameters, a-priori, on a contract basis: IR (ommitted Information Rate) [bit/s] S (ommitted urst Size) [bit] ES (Excess urst Size) [bit] IR: guaranteed bit rate (throughput) S: amount of data the network is willing to accept over a measurement period T ES: amount of excess data the network may transfer over T. Packets are marked with the E bit set to 1 ata exceeding S+ES are directly discarded at network access Network management and QoS provisioning - 5 Frame Relay: measurement interval T definition IR S ES T > 0 > 0 > 0 S/IR > 0 > 0 = 0 S/IR = 0 = 0 > 0 ES/ccess Rate Network management and QoS provisioning - 6 Pag. 2

3 Σ IR,j ESS_RTE where,j refers to the V from to j Network management and QoS provisioning - 7 Σ IR,j ESS_RTE where,j refers to the V from to j Network management and QoS provisioning - 8 Σ IR,j ESS_RTE where,j refers to the V from to j Network management and QoS provisioning - 9 Pag. 3

4 Σ IR,j ESS_RTE where,j refers to the V from to j Network management and QoS provisioning - 10 Σ IR,j ESS_RTE where,j refers to the V from to j Network management and QoS provisioning - 11 Σ LINK IR i,j LINK_SPEE links where i,j refers to the V from i to j Network management and QoS provisioning - 12 Pag. 4

5 Σ LINK IR i,j LINK_SPEE links where i,j refers to the V from i to j Network management and QoS provisioning - 13 Σ LINK IR i,j LINK_SPEE links where i,j refers to the V from i to j Network management and QoS provisioning - 14 Σ LINK IR i,j LINK_SPEE links where i,j refers to the V from i to j Network management and QoS provisioning - 15 Pag. 5

6 Frame Relay: congestion Summation of IR over all virtual circuits on each link may exceed the available bit rate over a link (overbooking) reates congestion, potentially a long-term congestion Traffic burstiness may create congestion (typically short term congestion) Need to control congestion? X.25 (ISN) may exploit link-by-link (hop-by-hop) flow control (and internal switch backpressure) to control (un-fairly) congestion In Internet the congestion control is delegated to hosts running TP, the network simply drops packets Frame relay, which does not implement flow control, uses explicit signaling from network nodes to signal congestion to users through FEN and EN bits Network management and QoS provisioning - 16 Frame Relay: algorithms Policing, or conformance verification Leaky ucket Token ucket ongestion control backward forward Network management and QoS provisioning - 17 onformance verification asic idea If a packet reaches network access and is conformant to the S constraint over T, it is transmitted at high priority with E=0 If a packet reaches network access and is not conformant to S over T but it is conformant to S+ES over T, it is transmitted at low priority with E=1 If a packet reaches network access and is not conformant to S+ES over T, it is discarded Same algorithms can be used to do shaping Traffic adaptation to make it conformant elay instead of marking/dropping Network management and QoS provisioning - 18 Pag. 6

7 Leaky ucket s a traffic regulator User traffic entering the buffer is transmitted at a maximum R rate equal to ρ User traffic exceeding the buffer size is dropped ny source becomes a R source at rate ρ If packet size is fixed When using to do conformance verification, if packets arrives earlier than they should be, with respect to ρ, drop it ρ Network management and QoS provisioning - 19 Leaky ucket uffer size is not a critical parameter: we can assume infinite buffer size mount of data sent over a period T is Tρ imensioning of parameter ρ for VR traffic ρ M : too much traffic could be discarded ρ P : waste of link bit rate, largely underutilized Traffic regulator which does not admit any burstiness Network management and QoS provisioning - 20 Token ucket ρ β Tk Token buffer bff source network Network management and QoS provisioning - 21 Pag. 7

8 Token ucket Tokens are generated at a fixed rate ρ maximum number of β tokens can be stored in the token buffer Permits some burstiness User data are sent over the network only if there is a token available in the token buffer Maximum amount of data send over a period T is Tρ + β The source becomes a VR source with M ρ P = access rate urst duration β ccess to the network can be further regulated with a cascading leaky bucket to limit P Network management and QoS provisioning - 22 Token ucket+ Leaky ucket Regulates average rate ρ 2, peak rate ρ 1, burst duration β ρ 2 β Token buffer source network ρ 1 Network management and QoS provisioning - 23 ll packets conformant to S bits iscard S+ES S ccess rate IR E=1 E=0 T 0 T 0 +T Frame arrival time Network management and QoS provisioning - 24 Pag. 8

9 One packet at low priority bits iscard S+ES S ccess rate IR E=1 E=0 T 0 T 0 +T Frame arrival time Network management and QoS provisioning - 25 One packet discarded bits iscard S+ES S ccess rate IR E=1 E=0 T 0 T 0 +T Frame arrival time Network management and QoS provisioning - 26 Measurement problems Measuring a rate of an asynchronous packet flow may be complex Simple solution: Measure over fixed length intervals Interval 1 Interval 2 time When does the interval starts? order effect between adjacent intervals? Network management and QoS provisioning - 27 Pag. 9

10 Measurements problems etter solution is a temporal sliding window of W seconds When a packet arrives, the rate is measured accounting for the amount of byte received during the last W seconds ifficult to implement (necessary to remember packet arrival times) It is possible to exploit a fluid approximation New rate R estimate at each packet arrival t packet arrival, we assume that the flow has transmitted R*W byte in the last W seconds and the estimate of R is updated Network management and QoS provisioning - 28 ongestion control Flow control is not supported in Frame Relay The network is unprotected against congestion Only protection mechanism is packet discarding ongestion should not occur if sources are sending at IR! When a switch (network node) establishes that congestion has occurred, to signal congestion it sets one among two bits: FEN (Forward technique) EN (ackward technique) Network management and QoS provisioning - 29 ongestion control: goals void packet loss onstraints? Maximum network utilization Fairness Often in contrast Simple case: all flows are alike Fairness means to provide the same set of resources to all flows Over a single bottleneck the problem is trivial Network wide problem Network management and QoS provisioning - 30 Pag. 10

11 ongestion control: an example 1 2 Link =10Mb/s 3 Link =10Mb/s Link =45 Mb/s Maximize the bit rate received by each flow Flow 1: 5 Mb/s Flow 2: 5 Mb/s Flow 3: 40 Mb/s Maximize the overall network utilization Flow 1: 10 Mb/s Flow 2: 0 Mb/s Flow 3: 45 Mb/s Network management and QoS provisioning - 31 Max-min fairness One possible definition of fairness bandwidth allocation is defined as max-min if It maximizes the bandwidth allocation to flows who receive the minimum allocation Property: max-min allocation is such that, to increase the bandwidth allocated to a flow, it is necessary to decrease the bandwidth allocated to another flow which is already a smaller or equal bandwidth In other words, no bandwidth increase can be obtained without penalizing flows already receiving a smaller allocation max-min allocation cannot be obtained with local assignments global network view is needed Network management and QoS provisioning - 32 Max-min fairness: algorithm Given: topology, link capacity, flows and flow routing 1) The algorithm starts with a 0 allocation to all flows, each flow is marked as unsatisfied 2) The allocation of all unsatisfied flows is increased by the same, small, quantity, until a bottleneck link is saturated 3) ll bottlenecked flows are saturated, thus, cannot receive a larger allocation ottlenecked flows are marked as satisfied 4) Goto 2, until all flows are bottlenecked and satisfied Must re-run for any topology or flow modification Network management and QoS provisioning - 33 Pag. 11

12 Max-min fairness: example Problem: find a fair bandwidth allocation to flows, according to the max-min paradigm S1 S Link 2 Link 4 R R R R Link 1 Link 3 R 100 Mb/s 34 Mb/s S3 S4 S5 S6 S7 S8 Network management and QoS provisioning - 34 Max-min fairness: example Order in which links are saturated L1, L4, L3, L2 Solution: fair max-min allocation S1: Mbps S2 : Mbps S3 : 17 Mbps S4 : 17 Mbps S5 : Mbps S6 : Mbps S7 : Mbps S8 : Mbps Network management and QoS provisioning - 35 Forward congestion When a switch detects congestion, it sets the FEN bit to 1 on all arriving packets sharing the congested buffer ongestion signaled to all congested Vs When the congestion indication reaches the receiver, it is redirected to the transmitter on a data flow traveling in the opposite direction The transmitter reduces the transmission speed according to a standardized algorithm Properties: Relatively slow Simple to implement No additional traffic is created, if there is a data flow from receiver to transmitter (normally at least Ks are sent) Network management and QoS provisioning - 36 Pag. 12

13 ackward congestion When a switch detects congestion, it sets the EN bit to 1 on all packets belonging to congested Vs These packets are not stored in the congested buffer! d-hoc signaling packets may be generated by the switch if no data traffic is flowing in the opposite direction The transmitter reduces the transmission speed according to a standardized algorithm when it detects packets with EN=1 Properties: Relatively fast omplex: need to store a list of congested LI on the forward path and to wait (or create after a timeout) packets with the proper LI on the backward path Network management and QoS provisioning - 37 Source behaviour: FEN The FEN technique is based on the idea that congestion phenomena are relatively slow Transmitter Start transmitting at a speed equal to IR omputes the percentage of LP-F frames received with a FEN bit set to 1 (FEN 1 ) over a pre-determined time interval If FEN 1 is >50%, the emission rate is reduced If FEN 1 is <50%, the emission rate is incremented Network management and QoS provisioning - 38 Source behaviour: FEN Measuring interval δ=2rtt Rate based transmitter: R INITIL = IR if FEN 1 > FEN 0 R New =0.875R Old if FEN 1 < FEN 0 R New=1.0625R Old If not transmitting for T, restart from R INITIL Window based transmitter: W INITIL = 1 if FEN 1 > FEN 0 W New =0.875W Old if FEN 1 < FEN O W New = W Old +1 Network management and QoS provisioning - 39 Pag. 13

14 Source behaviour: EN The EN technique is based on the idea that congestion phenomena are fast Instantaneous reaction (not based on δ=2rtt) R INITIL = IR If a single frame with EN=1 is received: R N =1/8R O If a single frame with EN=0 is received: Increase rate Network management and QoS provisioning - 40 ongestion control: issues Fairness among flows More active flows receive more congestion signals May be ok, since are creating more congestion, but is it max-min fair? Temporarily inactive flows? Signaling frequency ny reaction to congestion signals is constrained by the flow RTT It would make sense to adapt the (congestion) signaling frequency to flow RTT Practically impossible to know flow RTT in network nodes (may be done at the tx/rx side) onnections with shorter RTTs react faster oth when increasing and decreasing rate When congestion is detected (set up congestion bit in the header) Operate on packet reaching the buffer or leaving the buffer? Network management and QoS provisioning - 41 ongestion control: issues How to detect congestion? Measure ingress flow speed in each buffer Over which time interval? Worth complexity given the binary feedback available? lways balance complexity, performance, signaling capability Operates on a flow basis or on traffic aggregate? Threshold on buffer occupancy Instantaneous buffer occupancy Fast, but unstable Typically exploits some hysteresis to avoid switching between congested/non- congested state t verage occupancy over a time-sliding measurement window How the window size should be determined? More stable, but slower in reaction Occupancy derivative More precise than occupancy alone» uffer occupancy of 100 packets should be treated differently if the previous buffer occupancy was 150 or 50 packets Need to define time interval over which evaluate the derivative Threshold value? lose to zero occupancy to exploit most of the buffer size Enough space below threshold to allow for synchronous arrivals and avoid unneeded congestion signals Network management and QoS provisioning - 42 Pag. 14

15 ongestion control: issues uffer sizing? uffer above threshold should increase proportionally to Number of connections involved in congestion onnection RTTs Need to buffer in-flight packets onnection rate lways pay attention to The scenario in which algorithms are compared Network topology (often single bottleneck node examined) Number of flows Flow behavior ifficulty in properly setting parameters If choosing wrong values what happens? How difficult is to set up proper values? lgorithm robustness to parameter setting lgorithm complexity w.r.t. performance gain ll parameters (threshold, measurement window, buffer size) could be set off-line or modified at run time Run time modification is worth the effort? Network management and QoS provisioning - 43 Pag. 15