Abstract
Applications that require good network performanceoften use parallel TCP streams and TCP modificationsto improve the effectiveness of TCP. If the network bottleneckis fully utilized, this approach boosts throughput by unfairlystealing bandwidth from competing TCP streams. Improvingthe effectiveness of TCP is easy, but improving effectivenesswhile maintaining fairness is difficult. In this paper, we describean approach we implemented that uses a long virtual roundtrip time in combination with parallel TCP streams to improveeffectiveness on underutilized networks. Our approach prioritizesfairness at the expense of effectiveness when the network is fullyutilized. We compared our approach with standard parallel TCPover a wide-area network, and found that our approach preserveseffectiveness and is fairer to competing traffic than standardparallel TCP.Index Terms—System design, Network measurements, Experimentationwith real networks/Testbeds
I. INTRODUCTION
The ability to move large amounts of data quickly over ashared wide area network is a necessity for many applicationstoday. The Atlas project [1] must be able to move manypetabytes of data per year between Europe and the UnitedStates. Multiuser collaborative environments combine visualization,video conferencing, and remote application steeringinto a distributed application with low latency and high bandwidthdemands [2], [3]. The Optiputer project [4] aims to builda distributed computer using a wide area network path for abackplane. Other tools such as GridFTP [5], bbcp [6], DPSS[7], and PSockets [8] are used by applications that need tomove large amounts of data over wide area networks.Many of these applications use the Transmission ControlProtocol (TCP) for accurate and reliable in-order transmissionof data. TCP relies on the congestion avoidance algorithm(CAA) [9] to measure the capacity of a network path, fairlyshare bandwidth between competing TCP streams, and tomaximize the effective use of the network.Unfortunately, the CAA prevents the effective and efficientuse of wide area networks for these applications. The CAArelies on packet loss to indicate that the network is overloaded,and responds to packet loss by decreasing the transmission rateof the TCP stream by half. Van Jacobson [9] assumed threethings about packet loss when he designed the CAA – thefraction of non-congestion packet loss (”damaged packets”)is _ 1%; network speeds (circa 1988) are low enough toprevent measurable sensitivity to low (< 5%) loss rates; andan implicit assumption that packet reordering would not affectthe congestion avoidance algorithm. Jacobson found that theCAA is sensitive to packet loss when the packet loss rate isthe same order of magnitude as the square of the congestionwindow in packets. This has become a problem for high-speednetworks, since the maximum frame size is usually fixed at1500 bytes [10].There is a substantial body of evidence that effects suchas packet reordering and non-congestion packet loss candegrade TCP performance over wide area networks. The noncongestionlosses perceived by TCP are due to many factorsunrelated to simple network congestion that include end-host[11]–[15] and network infrastructure effects [16]–[19]. Theseeffects on wide-area high-speed networks can lead to situationsin which it can require over 4 hours for a TCP stream tocompletely recover from one loss event [20]. Non-congestionloss also has significant effects on satellite links [21]. There areefforts within the networking community to design modificationsto TCP congestion avoidance to overcome these effects.Most of these efforts rely on mechanisms that aggressivelycompete for network bandwidth, which can lead to unfairbehaviors at best, and congestion collapse at worst.In this paper, we describe an approach that modifies theaggressive tendencies of parallel TCP to allow an applicationto consume unused bandwidth when the network is underutilized.When the network is fully utilized, our approachis substantially fairer to competing traffic than an equivalentnumber of parallel TCP streams.
II. BACKGROUND
In previous work [22], we performed a series of networkmeasurements between the University of Michigan in AnnArbor and NASA AMES in Moffett Field, CA over a highspeed path to determine if parallel TCP streams were effectiveat improving throughput, and to gain insight on the relationshipbetween the number of parallel TCP streams and performance.We found that as the number of TCP sockets used in the parallel stream increased, aggregate throughput increased linearlyuntil the bandwidth ∗ delay product of the network wasachieved.We were convinced that parallel TCP was effective, but wereconcerned about the effects of parallel TCP on other networkstreams. To gain deeper insight into the effects of parallel TCP,we performed a series of simulations to measure the effects ofparallel TCP on effectiveness and fairness [23]. The simulationloss model was based on loss characteristics observed fromdata transfers over a wide area network. We found that thesimulation loss rate was high enough to prevent a single TCPstream from fully utilizing available network bandwidth, andthat a set of parallel TCP streams could collectively consumeunused bandwidth without stealing bandwidth from competingtraffic. However, when the network was fully utilized, parallelTCP could improve performance only by unfairly stealingbandwidth from competing traffic.In simulation, we discovered that a set of parallel TCPstreams that coupled one standard TCP stream with a groupof TCP streams with reduced aggressiveness could effectivelyconsume unused bandwidth, yet were unable to steal bandwidthfrom standard TCP streams when the network wasbusy. We found that we could adopt this method to retainthe desirable effectiveness of parallel TCP and substantiallyreduce unfairness when the network was busy.Our implementation modulates the aggressiveness of anindividual TCP stream by increasing the number of ACKsnecessary to increase the congestion window. This has theeffect of creating a virtual long round trip time (RTT) for thestream. Our approach exploits a behavior of TCP in whichlong RTT TCP streams are unable to compete with short RTTTCP streams for available network bandwidth [24]. We callthese modulated TCP streams fractional streams because theiraggressiveness is a fixed fraction of the aggressiveness of asingle unmodulated TCP stream. We call the combination ofthe unmodified single TCP stream and the set of fractionalstreams Combined Parallel TCP. This paper describes theimplementation and test of this modification on a real network.We believe that the work described in this paper contributesto existing work in several ways. First, our Combined ParallelTCP approach retains the TCP-friendly characteristics ofeach TCP stream. We exploit a behavior of TCP that doesnot fundamentally change the AIMD characteristics of TCP.Second, our work substantially retains effectiveness on underutilizednetworks and automatically prioritizes fairness overeffectiveness when the network is busy. Finally, our approachprovides a mechanism to dynamically tune the aggressivenessof parallel TCP using the virtual RTT factor.The next section of this paper describes other approachesused to improve TCP performance over wide area networks.III. RELATED WORKMany papers have been written on the topic of improvingTCP performance for long-lived TCP streams on high-speednetworks. Scalable TCP [20] improves TCP performance byusing a fixed additive increase (AI) value and a multiplicativedecrease (MD) factor that is less than the 1/2cwnd factorused by the standard TCP congestion avoidance algorithm.These AIMD factors make Scalable TCP more aggressivethan standard TCP by seizing more bandwidth while probingnetwork capacity and releasing less bandwidth on packet loss.Using these AIMD modifications, a Scalable TCP streamcompeting with a standard TCP stream on a shared networkpath with similar round trip times will not seek the optimalfairness and efficiency point in the phase diagram of Chiu andJain [25]. Thus, the use of competing Scalable TCP flows ona busy network will lead to a higher packet loss rates andreduce the amount of effective work (”goodput”) performedby the network.PSockets [8] is an application level library that opensconcurrent TCP streams to increase aggregate TCP throughput.PSocket exploits the fact that a set of N TCP streams isfunctionally equivalent to a virtual Maximum Segment Size(MSS) of N * MSS of the MSS used by a single TCP stream.Parallel TCP streams are aggressive on a shared network andcan steal bandwidth from competing TCP streams. Previouswork [22] showed that a single application using N parallelTCP streams competing with k other streams will receiveN/(N+k) of the network bandwidth, rather than the 1/(N+k)portion the other streams receive.High Speed TCP (HSTCP) [16] is designed to amelioratethe effects of media bit error rates on long-lived TCPstreams over wide-area networks. HSTCP modifies the AIMDparameters of the standard TCP congestion avoidance controlsystem as a function of the current cwnd and observed packetloss rate. This parameter modification is designed to gentlyprobe network capacity during the early startup phase of aHSTCP stream (when cwnd is small). After the gentle startupphase, when cwnd is large, HSTCP releases substantially lessnetwork bandwidth than 1/2cwnd on packet loss, and recoversmore aggressively than standard TCP during the AI phase.When the modified AIMD parameters are is use, HSTCP isunfair to competing unmodified TCP streams. Moreover, dueto its aggressive tendencies, HSTCP is unsuitable for use onnetworks slower than 10 Gb/sec, limiting its applicability.FAST TCP [26] uses both packet loss and increased packetround trip time as indications of network congestion. FASTTCP implicitly assumes that the minimum round trip timevalue observed during a TCP session is the propagationdelay (transmission delay with no queueing) on the endto-end network path. However, if FAST TCP is used onnetworks with persistently populated queues, it will incorrectlyinfer the propagation delay of the network and behave moreaggressively than competing unmodified TCP streams. Biaz[27] found that RTT measurements do not adequately representqueue length, and should not be used as a predictorof congestion. Measurements [26] show that FAST TCP isunfair to competing unmodified TCP streams using networkmeasurements from CERN to the United States over a 2.4Gb/sec network.


Download full report
http://citeseerx.ist.psu.edu/viewdoc/dow...1&type=pdf