spaper22_slides - A Network-Aware Approach for Video and...

Info iconThis preview shows page 1. Sign up to view the full content.

View Full Document Right Arrow Icon
This is the end of the preview. Sign up to access the rest of the document.

Unformatted text preview: A Network-Aware Approach for Video and Metadata Streaming Authors: Aravindan Raghuveer, Ewa Kusmierek, and David. H. C. Du Presented by: Terry Ching-Hsiang Hsu Martin Sarov Dec 1, 2011 Terry Ching-Hsiang Hsu, Martin Sarov 1 Dept. of CISE, University of Florida Talk Progress • • • • • • • Introduction Network Aware Rate Adaptation Dynamic Rate Adaptation Evaluation and Experiment Results Control Data Compression and Delivery Related Work Summary Dec 1, 2011 Terry Ching-Hsiang Hsu, Martin Sarov 2 Dept. of CISE, University of Florida What’s the Problem? • • • Internet-based video streaming applications! • Strict timing requirements, but Internet is best effort... Solutions? • YES! Network-aware demand adaptation Mechanism of network-aware demand adaptation: • • Evaluate the status of the network Perform adaptation based on this evaluation Dec 1, 2011 Terry Ching-Hsiang Hsu, Martin Sarov 3 Dept. of CISE, University of Florida Contributions • Dynamically adapt the bandwidth demand to the network conditions without always compromising on video quality • Compress the control information and thus reduce the delivery overhead • Propose DART - Dynamic Scheduling Algorithm for Reduced Trace delivery - to deliver the metadata to the client in a bandwidth-efficient and network-aware way. Dec 1, 2011 Terry Ching-Hsiang Hsu, Martin Sarov 4 Dept. of CISE, University of Florida Video and Metadata Unified Framework RAGHUVEER et al.: A NETWORK-AWARE APPROACH FOR VIDEO AND METADATA STREAMING 1029 Fig. 1. Unified architecture for video and metadata delivery. of control data ahead of playback and multiplex the remaining control data along the media stream. This paper makes three primary contributions. • We show that it is possible to dynamically adapt the bandwidth demand to the network conditions without always Dec 1, 2011 Terry Ching-Hsiang Hsu, Martin Sarov compromising on video quality. improving perceived quality. We use this finding as the underlying premise in this work. However in [5], only three rates are used and these rates are chosen in a static fashion without knowledge of the network conditions. In Section III, we point out the Dept. of selection of the sending 5 necessity of network-aware CISE, University of Floridarate and present a technique to choose it dynamically. Video and Metadata Unified Framework RAGHUVEER et al.: A NETWORK-AWARE APPROACH FOR VIDEO AND METADATA STREAMING A mismatch between demand and supply triggers an adaptation step 1029 Fig. 1. Unified architecture for video and metadata delivery. of control data ahead of playback and multiplex the remaining control data along the media stream. This paper makes three primary contributions. • We show that it is possible to dynamically adapt the bandwidth demand to the network conditions without always Dec 1, 2011 Terry Ching-Hsiang Hsu, Martin Sarov compromising on video quality. improving perceived quality. We use this finding as the underlying premise in this work. However in [5], only three rates are used and these rates are chosen in a static fashion without knowledge of the network conditions. In Section III, we point out the Dept. of selection of the sending 5 necessity of network-aware CISE, University of Floridarate and present a technique to choose it dynamically. Video and Metadata Unified Framework RAGHUVEER et al.: A NETWORK-AWARE APPROACH FOR VIDEO AND METADATA STREAMING A mismatch between demand and supply triggers an adaptation step 1029 Fig. 1. Unified architecture for video and metadata delivery. Server varies its sending rate to adapt to the network and client conditions of control contains the and multiplex the remaining Client data ahead of playbackadaptation logic. improving perceived quality. We use this finding as the under- control data along the media stream. This paper makes three primary contributions. • We show that it is possible to dynamically adapt the bandwidth demand to the network conditions without always Dec 1, 2011 Terry Ching-Hsiang Hsu, Martin Sarov compromising on video quality. lying premise in this work. However in [5], only three rates are used and these rates are chosen in a static fashion without knowledge of the network conditions. In Section III, we point out the Dept. of selection of the sending 5 necessity of network-aware CISE, University of Floridarate and present a technique to choose it dynamically. Talk Progress • • • • • • • Introduction Network Aware Rate Adaptation Dynamic Rate Adaptation Evaluation and Experiment Results Control Data Compression and Delivery Related Work Summary Dec 1, 2011 Terry Ching-Hsiang Hsu, Martin Sarov 6 Dept. of CISE, University of Florida Network Aware Rate Adaptation • • • Network status evaluation 3Rate adaptation scheme Assumption: clients should have fairly accurate information about the network status. Dec 1, 2011 Terry Ching-Hsiang Hsu, Martin Sarov 7 Dept. of CISE, University of Florida Network Status Evaluation • • Available bandwidth (measured at client) • Ra: arrival rate, Rs server sending rate, Ba: available bandwidth • • If Ra < Rs, then Ba < Rs If Ra ≥ Rs, then Ba ≥ Rs Network backlog (via MECN capable routers) • • MECN: multiple ECN Queue length thresholds: min_th, mid_th, max_th Dec 1, 2011 Terry Ching-Hsiang Hsu, Martin Sarov 8 Dept. of CISE, University of Florida 3Rate Adaptation Scheme • • Arrival rate is more fluctuating compared to sending rate. • Client buffer is used to cushion the fluctuation. Client buffer starvation leads to switching to low quality level with lower bandwidth requirements. • • Solution: 3Rate scheme increase server’s sending rate hoping that the arrival rate would consequently increase. • • Increase sending rate: no congestion Quality drop: congestion detected Default rate r, low rate rl, high rate rh. Dec 1, 2011 Terry Ching-Hsiang Hsu, Martin Sarov 9 Dept. of CISE, University of Florida 3Rate Adaptation Scheme (cont.) 1030 IEEE TRANSACTIONS ON CIRCUITS TABLE I 3RATE ADAPTATION SCHEME starvation may occur and the client is forced to switch to a Dec 1, 2011 Terry Ching-Hsiang Hsu, Martin Sarov Dept. of CISE, University of Florida 10 lower quality level with lower bandwidth requirements. The A. A A flow prop in tw will that is a of th whe imu Downside of 3Rate Adaptation Scheme • Switching to statically chosen sending rates rh and rl may not be optimal. • Predetermined rh not high enough → arrival rate may not be sufficient to avoid underflow. • Predetermined rl not low enough → client is too aggressive in acquiring bandwidth • Observation: changing sending rate without taking the network conditions and client buffer occupancy into consideration may lead to suboptimal results! Dec 1, 2011 Terry Ching-Hsiang Hsu, Martin Sarov 11 Dept. of CISE, University of Florida Talk Progress • • • • • • • Introduction Network Aware Rate Adaptation Dynamic Rate Adaptation Evaluation and Experiment Results Control Data Compression and Delivery Related Work Summary Dec 1, 2011 Terry Ching-Hsiang Hsu, Martin Sarov 12 Dept. of CISE, University of Florida Dynamic Rate Adaptation • • • • • Algorithm for dynamic rate increase Algorithm for dynamic rate decrease Algorithm for quality increase Dynamic rate, quality adaptation TCP friendliness of D-rate Dec 1, 2011 Terry Ching-Hsiang Hsu, Martin Sarov 13 Dept. of CISE, University of Florida ATION SCHEME A rate increase is done when the client faces a buffer underflow and there is no incipient congestion in the network. The proposed algorithm dynamically calculates the new sending rate in two steps. We first find a target arrival rate at the client, which will avoid the buffer underflow. Then we find transmission rate that will result in the targeted arrival rate. The new sending rate D SYSTEMS FOR VIDEO TECHNOLOGY, VOL. 17, NO. 8, AUGUST 2007 Rate increase only when the targeted arrival rate and the responsiveness is a function of of the network and is computed as Algorithm for Dynamic Rate Increase • Buffer underflow, • Dynamic Rate Increase and rithm for e increase isNo incipient client faces a buffer underthe congestion in the network • nodone whencongestion in the network. The d there is incipient Dynamically d algorithm dynamically calculates thethe new sending rate in two steps sending rate • first find a targetcalculatesat newclient, which teps. We arrival rate the client id the buffer underflow. Then we findrate at the rate •theFind a target arrival transmission rate to avoid underflow l result in targeted arrival rate. The new sending ction of • Find transmission rate responsiveness the targeted arrival rate and the that will result in the targeted (1) client is forced to switch to a where is the network response index and is the minbandwidth requirements. The imum arrival rate required to avoid buffer underflow. is the e, and possibly avoid, such a safety factor to make sure that we do slightly server’s sending rate hoping better than the minimum. Network response index is calculated equently increase. However, the as , where is the current transmission rate and n there is no congestion in the is the corresponding arrival rate. represents the minimum ested when there is congestion. transmission rate required for the arrival rate to reach value re three sending rates: the high . The calculations are done at the granularity of a single the low rate . The client can frame playback time in a discrete time model adopted. Specifirary switch to arrival etwork and either orrate cally, the minimum arrival rate required is computed as is computedinas se the arrival rate. The network ancy are periodically evaluated, (1) (2) , tion is done as shown in Table I. and high ermark is the time at which is the network watermark index and response where is the minbuffer underflow and overflow, the rate increase step is initiated and represents the size of rrival rate required to avoid buffer underflow. is the the predicted buffer occupancy the frame that is played out at time . Let be the value of for Dec Hsu, 14 actor1, 2011 Terry Ching-Hsiangcal- Martin Sarov maximum is slightly Dept. of CISE, University of Florida e round-trip time (rtt). It is to make sure that we do reached in (2). The length of the time which the • Algorithm: Afor Dynamic RateVIDEO AND METADATA STR Decrease RAGHUVEER et al. NETWORK-AWARE APPROACH FOR By sending at higher the we when the more backlog •handle. This dataa reaches rate,clientcould sendnetworkdata into the network than what the client buffer can handle... bad! clears up causing buffer overflow. To handle this situation, the rate decrease step is used. The client keeps track of thesent Client keeps track of the additional amount of data addias compared rate default transtional amount of datato the default sending to ther. ∆D as compared sent mission schedule with sending rate . The new sending rate is The new sending rate is calculated as calculated as • • (3) • and ∆tdec iis the time for which rate decrease is is requested. s the for which rate decrease requested. The is bounded by the predicted time to maximum value of for which overflow, calculated Sarov the maximum as Dec 1, 2011 Terry Ching-Hsiang Hsu, Martin Dept. of CISE, University of Florida 15 con Algorithm for Quality Increase Bandwidth probing! Increase sending rate by a small step Congestion? No No Yes Yes Suspend experiment and remove extra data Dec 1, 2011 Terry Ching-Hsiang Hsu, Martin Sarov sending rate reaches higher quality video? 16 Quality increase Dept. of CISE, University of Florida Algorithm for Quality Increase Bandwidth probing! Increase sending rate by a small step Congestion? No No sending rate reaches higher quality video? Yes Yes Suspend experiment and remove extra data Quality increase Bandwidth probing phase is initiated only when there the client has extra buffer to support the experiment Dec 1, 2011 Terry Ching-Hsiang Hsu, Martin Sarov 16 Dept. of CISE, University of Florida 7: return threshold) packet is severe to interpret (third network backlog. Rate used congestion the frames prompts a quality drop. increase cannot be performed if the first queue length b) Arrival Rate: The current arrival rate is calculated every 1) least router. An indication of threshold is reached in at Network Resource Availability Metrics: (4) seconds as where is the packet built, severe congestion predicted arrival rateBacklog:a TheisECN value in each incom a) Network for the calcusize. The (third threshold) promptsnext quality que drop. to lated using the exponential smoothing technique. The pa- network backlog. R packet is used to interpret the higher the weight is calculated rameter controlsincrease given ) Arrival Rate: The current arrival ratecannot be performed if the first queue len ytes) at Network resource availability to historyevery time metrics: to this seconds as where is the packet se thethat is played threshold is reached in at least router. An indicatio ame calcu- (5) size. The predicted arrival rate for the value iniseach packet. next Network backlog: E congestion ffer. using the exponentialsevereCN technique. (third threshold) prompts a qua lated smoothing The paate in2) Client Metric: Buffer Occupancy: The buffer occupancy drop. controlsrate: current based on the is rate. The rameterpredicted 2rttthe weight given to history arrivalcalculated every ∆t band- is Arrival into the future arrival rate Dynamic Rate (D-Rate), Quality Adaptation • • • occupancy after Arrival Rate:the current time arrival rate is calculated ev b) seconds from The current e, the buffer is (5) seconds as nd the calculated by adding to the current buffer ,occupancy, the differ• between number ofsize. The predicted arrival rate forwherenext is the pa has ence cation been built, frames expected to be received and the is ca the One number of Buffer that will be consumed in obser- technique to 2) ClientClient metric: buffer occupancy occupancy smoothing technique. The Metric: frames Occupancy: The buffer lated • the higherthe future based using the exponential r than hift to predicted 2rtt into on the arrival rate. The rameter controls the weight given to history wever, is ffer occupancy after We refer to this seconds from the current time • adding to the current buffer occupancy, the differexperculated by tead increase the a rate ce between number of frames expected to be received and the (6) only if mber of frames that will be consumed in e conduct rate∆t ≤2rtt, k: highest frame index in the buffer, fps: consumption rate. next • 0≤ in2) Client Metric: Bufferin the buffer, where is the highest frame index Occupancy: The buffer occupa found h available band- consumption rate.2rtt into the futurethe un- on the arrival rate. is the is predicted The function uses based and derflow the buffer Sarov nDec 1, 2011 Terry Ching-Hsiang Hsu, Martin occupancy after sizes) toept. offrom the current time this phase, curve (cumulative function of frame seconds CISE, University of Florida D trans17 dvanlate the predicted amount of data that will arrive to number ConvertToFrames(data) EO AND METADATA STREAMING • Pseudocode of the function ξ() work backlog situation, the k of the addidefault transending rate is 1031 convertToFrames 1: 2: while counter < data do 3: (3) equested. The dicted time to ch 5: 6: end while 7: return frames 1) Network Resource Availability Metrics: a) Network Backlog: The ECN value in Deach CISE, University of Florida incoming Terry Ching-Hsiang Hsu, Martin Sarov ept. of 18 packet is used to interpret the network backlog. Rate (4) Dec 1, 2011 4: The Adaptation Algorithm (D-Rate) • • • • Client measures rtt and arrival rate periodically. Buffer occupancy after 2rtt is predicted. If buffer underflow and no congestion observed, • client requests rate change to rnew If quality drop has to be done, • quality is dropped based on the severity of congestion indicated by the ECN value. Dec 1, 2011 Terry Ching-Hsiang Hsu, Martin Sarov 19 Dept. of CISE, University of Florida TCP Friendliness of D-Rate • • Controlled rate increased steps • • • Magnitude: maximum rate is bounded by Rmax Duration: requests rate increase for a bounded duration Frequency: carefully monitor rate increases requests Pre-emptive quality drop • • Rate increase is employed only when no congestion When there is signs of congestion, D-Rate requests for quality drop in anticipation of a congestion. Dec 1, 2011 Terry Ching-Hsiang Hsu, Martin Sarov 20 Dept. of CISE, University of Florida Talk Progress • • • • • • • Introduction Network Aware Rate Adaptation Dynamic Rate Adaptation Evaluation and Experiment Results Control Data Compression and Delivery Related Work Summary Dec 1, 2011 Terry Ching-Hsiang Hsu, Martin Sarov 21 Dept. of CISE, University of Florida Evaluation and Experimental Results: Background • Goal: • Illustrate features of D-Rate • Compare and contrast adaptation schemes • • • • D-Rate 3Rate Pure Quality (PQ) Simulation Testbed: Video Server and Client: UDP and TCP connection MECN capable routers: 10 intervening nodes Cross traffic generators: a set of CBR traffic sources • • • Dec 1, 2011 Terry Ching-Hsiang Hsu, Martin Sarov 22 Dept. of CISE, University of Florida Evaluation and Experimental Results: Scenario 1 • Goal: DemonstrateTRANSACTIONS ON CIRCUITS AND SYSTEMSaFOR VIDEO TECHNOLOGY, VOL. 17, NO. 8, AUGUST 2007 IEEE the network awareness nd dynamic rate selection properties of D-Rate • riendliness of D-Rate • D maintaining constant video quality iendliness and etails: nflicting, yet important, demands that Internet-based aming applications should address. We highlight in Star Wars MPEG-4 encoded ing paragraphs some salient features of D-Rate that video traceout that there is scope P-friendly. However we point ement of the D-Rate scheme on the TCP friendliness Rate of to be added). . a slow start phase needs cross traffic is trolled Rate Increase Steps: between 55s and increased A rogue rate increase ate can lead to unfair consumption of bandwidth and 70s so that arrival rate at mpeting TCP applications. But D-Rate addresses this controlling the rate increaseis greater than each hop in three fronts: magnition, and frequency. First of all, the maximum rate the linka capacity.ned parameter requested is bounded by system defi ch is adjusted based on the history/knowledge of the Fig. 2. Star Wars: Underflow curve and transmission plan. m link characterstics or equation-based TCP friendDec [11]. Terry Ching-Hsiang Hsu, Martin Sarov Dept. of CISE, University of Florida 23 niques 1, 2011Next, we always request a rate increase ded duration, set in our experiments to 1 second. Fi- • • res of D-Rate that t that there is scope e TCP friendliness dded). ogue rate increase n of bandwidth and Rate addresses this hree fronts: magnithe maximum rate defined parameter y/knowledge of the based TCP frienduest a rate increase nts to 1 second. Fietwork is carefully his ensures that the rol. ncrease step is emgestion in the netas observed over a a quality drop (to pation of a congeskes a good balance e time maintaining Evaluation and Experimental Results: Scenario 1 (cont.) • Results: • Backlog builds up but does not exceed min_th Fig. 2. Star Wars: Underflow curve and transmission plan. Fig. 3. Scenario-1 (without adaptation): Average Dec 1, 2011 Terry Ching-Hsiang Hsu, Martin Sarov 24 lts to illustrate the sizes. Dept. of CISE, University of Florida Evaluation and Experimental Results: Scenario 1 (cont.) • Results: • Arrival rate < Sending rate due to backlog RAGHUVEER et al.: A NETWORK-AWARE APPROACH FOR VIDEO AND METADATA STREAMING Dec 1, 2011 Terry Ching-Hsiang Hsu, Martin (without adaptation): Arrival rate. 25 Fig. 4. Scenario-1 Sarov Dept. of CISE, University of Florida Fig. 7. Scenario-1: Arri Evaluation and Experimental Results: Scenario 1 (cont.) Fig. 4. Scenario-1 (without adaptation): Arrival rate. • • No congestion • • ra q PQ requests quality drop T a a c to n 3Rate, D-Rate request rate increase Effects on client buffer occupancy: • • • Fig. 5. Scenario-1: Buffer Occupancy. PQ: no underflow b/c reduced quality -> less required bandwidth 3Rate: underflow present b/c requested high rate is not sufficient D-Rate: no underflow Dec 1, 2011 Terry Ching-Hsiang Hsu, Martin Sarov F 26 Dept. of CISE, University of Florida s th te in th b g ra w s s w w Evaluation and Experimental Results: Scenario 1 (cont.) rate (and consequently arrival rate) at 70 s, corresponds to the quality drop requested. 3) Scenario-2: In the next set of experiments, we analyse the TCP friendliness of the D-Rate scheme during the rate increase and quality decrease steps. We attach a TCP-Vegas source (a ftp application) to the video server node and a TCP sink to the video client node. We increase the bandwidth of links by 0.422 Mbps to ensure that the TCP flow gets a fair share of bandwidth during no-congestion periods. First, we repeat the experiment in Scenario-I with the new setup and observe the throughput of the TCP flow. We observe that the buffer occupancy, arrival rate and sending rate charactersitcs of the video stream are very similar what was observed VEER et al.: A NETWORK-AWARE APPROACH FOR VIDEO AND METADATA STREAMING 1033 in Scenario I. Fig. 9 shows the throughput and window size of Fig. 5. Scenario-1: Buffer Occupancy. the TCP flow. We see that although the TCP throughput drops by a small amount during the rate increase step (62–69 s), it regains it pretty quickly soon after. Also since D-Rate increases rate only when there is no evidence of congestion in the network, it does not cause a reduction in TCP’s window size (as seen in the upper part of Fig. 9). In the next experiment, we increase the rate of the cross traffic so that the first queue length threshold is reached When faced with a possible underflow and signs of congestion in the network, D-Rate drops the quality by 0.5 (Fig. 10). The TCP sender also throttles its throughput rapidly and regains equally quick. D-Rate tries to smoothen out the rate changes so as to not affect the quality of the video stream adversely. 4) Scenario-3: In this experiment, we further illustrate the behavior of the D-Rate adaptation algorithm with higher levels of backlog at the core routers. We simulate network and client Scenario-12011 Sending Rate.Arrival Hsu, Fig.DecScenario-1:Terry Ching-Hsiangrate. Martin Sarov 6. 1, (without adaptation): D 27Scenario-1: Arrival rate. ept. to cases 3, 4, and of Florida Fig. 7. conditions corresponding of CISE, University 6 in Table I. We use the video trace of The Firm in this experiment. The underflow • Arrival rate for D-Rate is higher than rate for 3Rate b/c of higher sending rate selection • Sending/Arrival rate drop for PQ b/c of quality drop request Evaluation and Experimental Results: Scenario 2 • Goal: • IEEE riendliness of D-Rate during the rate increase and Analyze the TCP fTRANSACTIONS ON CIRCUITS AND SYSTEMS FOR VIDEO TECHNOLOGY, VOL. 17, NO. 8, AUGUST 2007 quality decrease steps CP Friendliness of D-Rate • Details: P Friendliness and maintaining constant video quality o conflicting, yet important, demands that Internet-based streaming applications a TCP-VegasWe highlight in Attach should address. source to llowing paragraphs some salient features of D-Rate that video server node and a is scope it TCP-friendly. However we point out that there TCP sink provement of o the client node TCP friendliness tthe D-Rate scheme on the ts (eg. a slow start phase needs to be added). Controlled Rate Increase Steps: A rogue rate increase Increase bandwidth of links by n D-Rate can lead to unfair consumption of bandwidth and ng competing0.422 Mbps to provide the TCP TCP applications. But D-Rate addresses this rn by controlling thewith a fairin three fronts: magniflow rate increase bandwidth share duration, and frequency. First of all, the maximum rate d is bounded by a system defined parameter an be requesteduring no-congestion periods. , which is adjusted based on the history/knowledge of the Fig. 2. Star Wars: Underflow curve and transmission plan. stream link characterstics or equation-based TCP friendDec 1, 2011 Terry Ching-Hsiang Hsu, Martin a rate Dept. of CISE, University of Florida techniques [11]. Next, we always request Sarov increase 28 bounded duration, set in our experiments to 1 second. Fi- • • Evaluation and Experimental Results: Scenario 2 • Part1: Fig. 8. The Firm: Underflow curve and transmission plan. • • Fig. 11 repeat experiment in Scenario 1 with the new setup Results (Pt1): • Buffer occupancy and arrival/sending rate ow with rate increase. Fig. 9. Scenario-2: TCP fl characteristics are similar to ones in Scenario 1. • TCP throughput drops a little during rate increase step but it regains quickly soon after • No reduction in TCP’s window size since D-Rate increases rate only when there is no congestion Dec 1, 2011 Terry Ching-Hsiang Hsu, Martin Sarov 29 Dept. of CISE, University of Florida Fig. 12 buffer not pe In is pus thresh ever th to the only a chang tion b Evaluation and Experimental Results: Scenario 2 Fig. 9. Scenario-2: TCP flow with rate increase. • • Part 2: • increase cross traffic rate so that first queue length threshold is reached Results (Pt2): • Fig. 10. Scenario-2: TCP flow with quality drop. Due to possible underflow and signs of congestion D-Rate drops quality by 0.5 because no information is available on how long the network TCP sender throttles throughput last. backlog may rapidly and regains it equally In the second phase (time period 75–95 s), the rate of the cross quickly. but traffic is increased further to push the backlog above . As a result the ECN bits of below the middle threshold Dec 1, 2011 Terry Ching-Hsiang Hsu, Martin Sarov packets are30 D an of CISE, University of Florida marked to indicate ept.incipient congestion. As the • Fig. 12 buffe not p In is pu thresh ever t to the only chang tion b Th D-Ra of ph reque ness i that a The p (Phas 5) the qu trace erage requi erage Evaluation and Experimental Results: Scenario 3 • Goal: 1034 • • IEEE TRANSACTIONS ON CIR Further illustrate behavior of D-Rate with higher levels of backlog Details: • • The Firm video trace Fig. p The Firm Underflow to: Adjust cross traffic rate in three 8. hases :aiming curve and transmission plan. • • • (0-75s) keep backlog below min_th (75-95s) push backlog between min_th and mid_th (110-117s) push backlog between mid_th and max_th Dec 1, 2011 Terry Ching-Hsiang Hsu, Martin Sarov 31 Dept. of CISE, University of Florida Evaluation and Experimental Results: Scenario 3 034 IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS FOR VIDEO TECHNOLOGY, VOL. 17, NO. 8, AUGUST 2007 • Results: • Phase 1: Buffer underflow is predicted and sending rate is increased • Phase 2: Buffer occupancy is healthy so no adaptation; The Firm: Underflow curve and transmission plan. ECN bits marked to indicate incipient congestion . 8. The Firm: Underflow curve and transmission plan. ig. 8. • Fig. 11. Scenario-3: Sending rate and ECN (D-Rate). Fig. 11. Scenario-3: Sending rate and ECN (D-Rate). Phase 3: Client request a quality drop; ECN bits marked to indicate moderate congestion Dec 1, 2011 Terry Ching-Hsiang Hsu, Martin Sarov . 9. Scenario-2: TCP flow with rate increase. 32 Dept. of CISE, University of Florida Fig. 12. Scenario-3: Buffer occupancy (D-Rate). Evaluation and Experimental Results: Scenario 4 • Goal: RAGHUVEER et al.: A NETWORK-AWARE APPROACH FOR VIDEO AND METADATA ST • • Demonstrate the quality increase algorithm in D-Rate Details: • • • • The Firm video trace Fig. stream < bandwidth required for Bandwidth available to video 13. Scenario-4b: Quality increase experiment. highest quality level the 40s after sufficient no greater than bandwidth First quality increase starts atserver-client path is set to a value congestion requirement of the highest quality level. history The rate of the cross traffic is chosen such that the bandwidth available to the video stream is lesser than bandwidth requireSecond quality increase starts of the highest quality level. The available bandwidth is ment at 80s but is suspended during bandwidth probing b/c min_th issufficient to service the previous quality level. however reached Quality increase experiments are started twice (at 40 and 80s) Dec 1, 2011 Terry Ching-Hsiang Hsu, Martin Sarov after building sufficient Dept. of CISE, University ofThe increment no congestion history. Florida 33 amount used was 0.025 Mbps. Fig. 13 shows the ECN • A. the pro tio sen an ap tio ex err po wh co Talk Progress • • • • • • • Introduction Network Aware Rate Adaptation Dynamic Rate Adaptation Evaluation and Experiment Results Control Data Compression and Delivery Related Work Summary Dec 1, 2011 Terry Ching-Hsiang Hsu, Martin Sarov 34 Dept. of CISE, University of Florida Control Data Compression and Delivery • Main concept: Underflow curve – cumulative function of frame sizes with respect to their play-out times • Key point: D-Rate requires the underflow curve (control data) of video stream • Two challenges: • Overhead due to sending this metadata could be substantial • Control data (like media data) is time-sensitive Dec 1, 2011 Terry Ching-Hsiang Hsu, Martin Sarov 35 Dept. of CISE, University of Florida • Control Data Compression and Delivery (cont.) Solutions (in brief): • • Problem 1: compress control data by approximating the underflow function with a piece-wise linear one. Each piece is a (Np, Dp) pair • • Np is number of frames in portion p Dp is amount of data contained in portion p Problem 2: use a delivery algorithm (DART) which: • • • • Delivers all (Np, Dp) pairs before their deadlines treq(p) Minimizes startup delay incurred by metadata delivery Adapts (dynamically) to network conditions Controls/reduces bandwidth usage of metadata delivery Dec 1, 2011 Terry Ching-Hsiang Hsu, Martin Sarov 36 Dept. of CISE, University of Florida Talk Progress • • • • • • • Introduction Network Aware Rate Adaptation Dynamic Rate Adaptation Evaluation and Experiment Results Control Data Compression and Delivery Related Work Summary Dec 1, 2011 Terry Ching-Hsiang Hsu, Martin Sarov 37 Dept. of CISE, University of Florida Related Work • Network aware adaptation techniques Receiver driven: clients modify its bandwidth demands based on network status. • Sender driven: sender uses feedback reports from the receiver to learn network status and adapt the outgoing bandwidth for that client. • • • Transcoder driven: gateways are placed at strategic points in network to vary the quality based on network status in each region Control data reduction and delivery schemes • MCBA, a bandwidth smoothing algorithm Dec 1, 2011 Terry Ching-Hsiang Hsu, Martin Sarov 38 Dept. of CISE, University of Florida Talk Progress • • • • • • • Introduction Network Aware Rate Adaptation Dynamic Rate Adaptation Evaluation and Experiment Results Control Data Compression and Delivery Related Work Summary Dec 1, 2011 Terry Ching-Hsiang Hsu, Martin Sarov 39 Dept. of CISE, University of Florida Summary • • • A Network aware video streaming technique that adapts sending rate of the video to match • • • time-varying network characteristics client data requirements buffer occupancy A a technique efficiently deliver metadata, synchronized with the media stream. Simulation shows the proposed techniques can efficiently adapt to changes in the network to provide better QoS. Dec 1, 2011 Terry Ching-Hsiang Hsu, Martin Sarov 40 Dept. of CISE, University of Florida Questions? Dec 1, 2011 Terry Ching-Hsiang Hsu, Martin Sarov 41 Dept. of CISE, University of Florida ...
View Full Document

Ask a homework question - tutors are online