On Parameter Tuning of Data Transfer Protocol GridFTP for Wide-Area etworks Takeshi Ito, Hiroyuki Ohsaki, and Makoto Imase Abstract In wide-area Grid comuting, geograhically distributed comutational resources are connected for enabling efficient and large-scale scientific/engineering comutations. In the wide-area Grid comuting, a data transfer rotocol called GridFTP has been commonly used for large file transfers. GridFTP has the following features for solving roblems of the existing TCP. First, for accelerating the start-u in TCP s slow start hase and achieving high throughut in TCP s congestion avoidance hase, multile TCP connections can be established in arallel. Second, according to the bandwidth-delay roduct of a network, the TCP socket buffer size can be negotiated between GridFTP server and client. However, in the literature, sufficient investigation has not been erformed either on the otimal number of TCP connections or the otimal TCP socket buffer size. In this aer, we therefore quantitatively investigate the otimal arameter configuration of GridFTP in terms of the number of TCP connections and the TCP socket buffer size. We first derive erformance metrics of GridFTP in steady state (i.e., goodut and acket loss robability). We then derive the otimal arameter configuration for GridFTP and quantitatively show erformance limitations of GridFTP through several numerical examles. We also demonstrate validity of our aroximate by comaring results with analytic ones. Keywords Grid Comuting, GridFTP, TCP (Transmission Control Protocol), Parameter Tuning, Steady State Analysis, Parallel TCP Connection, TCP Socket Buffer I. ITODUCTIO TASMISSIO Control Protocol (TCP) has been widely used as a transort-level communication rotocol in the Internet [1]. GridFTP has been designed for utilizing TCP as its transort-level communication rotocol [2]. However, TCP is a rather old communication rotocol that was designed in the 197s. Several roblems have been reorted regarding TCP such as its inability to suort the raidly increasing seeds of recent networks. As an examle, the current TCP eweno (TCP version eweno) cannot detect congestion in a network until acket loss occurs, so a large number of ackets are lost. With the faster seeds of networks and larger buffer sizes of routers in a network, the amount of ackets lost mushrooms and TCP throughut deteriorates significantly. To resolve existing TCP roblems, GridFTP has features such as establishing multile TCP connections in arallel to accelerate start-u in the TCP slow start hase and negotiating the TCP socket buffer size between the GridFTP server and client according to the bandwidth-delay roduct of a network [2]. However, the effectiveness of these features has not been fully investigated. In other words, otimal configurations for T. Ito, H. Ohsaki, and M. Imase are with Graduate School of Information Science and Technology, Osaka University. e-mail: {t-itou, oosaki, imase}@ist.osaka-u.ac.j the number of arallel TCP connections and TCP socket buffer size have not been investigated. There have been several related works on TCP socket buffer size and arallel TCP connections [3] [9]. In [3], [4], automatic tuning mechanisms of TCP socket buffer have been roosed. However, both aroaches require some modifications to a socket API and/or a TCP rotocol stack. In Grid comuting, heterogeneous comuting resources are integrated. Hence, such modifications to oerating systems are unrealistic; i.e., otimization of TCP socket buffer size should rely on information obtained from Grid middleware. In [5], an extension to GridFTP rotocol for automatically negotiating TCP socket buffer size has been roosed. However, the roosed mechanism is simle and not otimal; i.e., it simly allocates twice of the BDP (bandwidth-delay roduct) for each TCP connection. It is known that a larger TCP socket buffer does not always result in better erformance when acket loss rate is high [1]. On the contrary, effectiveness of arallel TCP connections has been studied by many researchers [6] [9]. In [6] [9], erformance of arallel TCP connections is investigated using exeriments. However, for otimizing the number of arallel TCP connections, -based aroaches are inaroriate; i.e., it is quite difficult or, in most cases, imossible to aly results for a arameter otimization. For otimizing the number of arallel TCP connections, some insight in the effect of arallel TCP connections on their erformance is necessary. In [7], [9], simle analytic models of arallel TCP connections have been resented. However, those analytic models are not alicable for otimizing the number of arallel TCP connections since they do not cature the trade-offs in arallel data transfer, as we will exlain in Section II. In this aer, by articularly focusing on the number of arallel TCP connections and TCP socket buffer size, we quantitatively investigate their otimal arameter configurations and clarify erformance limitations of GridFTP. We first derive a continuous-time model for GridFTP by aggregating multile TCP continuous-time models [11]. Since TCP is a feedback control that changes the window size deending on acket loss robability in a network, we model multile TCP connections as indeendent continuous-time SISO (Single-Inut and Single-Outut) systems. By combining these continuoustime TCP models, we then obtain a continuous-time model of GridFTP. Performance metrics in steady state (e.g., GridFTP goodut and acket loss robability) are derived using our GridFTP model. By focusing on the number of TCP connections and TCP socket buffer size, we also derive the otimal
arameter configuration for GridFTP and quantitatively show erformance limitations of GridFTP. ote that GridFTP suorts two tyes of using arallel TCP connections: arallel data transfer and stried data transfer (see Section II). The arallel data transfer (from a single server to a client) is a secial case of the stried data transfer (from multile servers to a client). For brevity, we limit the scoe of this aer to the arallel data transfer. However, our analytic aroach can be directly alied to the case of the stried data transfer by simly differentiating round-tri times,, in Section IV. Using a simle model enables us to derive several GridFTP erformance metrics in a closed-form, which gives us more insights than with an unnecessarily comlicated model. The structure of this aer is as follows. First, Section II briefly exlains GridFTP s major features (i.e., autonegotiation of TCP socket buffer size and arallel data transfer), and discusses unresolved roblems of GridFTP. Section III exlains the definition of terms used in this aer and our aroach for modeling GridFTP. Section IV analyzes steady state erformance of GridFTP by aggregating multile TCP continuous-time models, and derives the otimal arameter configuration of GridFTP. Section V comares analytic results with ones for validating our aroximate. Finally, Section VI summarizes this aer and discusses future toics. II. GIDFTP Standardization of GridFTP [2] as a common bulk data transfer rotocol for the Grid has been currently roceeding in the Global Grid Forum (GGF) [12]. GridFTP is a rotocol that extends FTP (File Transfer Protocol) [13] [15], which was standardized in the IETF and has been widely used in the Internet. In addition to features of the original FTP, the following features are added to GridFTP: auto-negotiation of TCP socket buffer size, arallel data transfer, third-arty control of data transfer, artial file transfer, security, and suort for reliable data transfer. In what follows, an overview of two major features of GridFTP, auto-negotiation of TCP socket buffer size and arallel data transfer, and unresolved roblems of GridFTP is resented. A. Auto-egotiation of TCP Socket Buffer Size In GridFTP, the server s TCP socket buffer size can be exlicitly configured by the client with the SBUF (Set Buffer Size) command. In addition, TCP socket buffer size can be configured by negotiating between the GridFTP server and client using the ABUF (Auto-egotiate Buffer Size) command. Almost all existing TCP imlementations allocate a fixedsize (e.g., 64 [Kbyte]) TCP socket buffer, so throughut imrovement can be exected when the TCP socket buffer size is aroriately configured according to the bandwidth-delay roduct of the network. However, it has not been adequately studied how the ABUF command should be imlemented. For instance, the ABUF command has not been imlemented in the GridFTP imlementation included in a Globus Toolkit [16]. Examles of the ABUF command imlementation is discussed in [2], [5], which measure a round-tri time and available bandwidth of a network by generating measurement traffic between the GridFTP server and client. However, more investigation on the ABUF command imlementation is necessary. In a real network, the available bandwidth of a network varies with time, and a large amount of measurement traffic must be generated for accurately measuring the available bandwidth. Because of these reasons, active measurement aroaches as discussed in [2], [5], which configure TCP socket buffer size based simly on the measured bandwidth-delay roduct, are inadequate for ractical uroses. amely, in a real network, otimization of GridFTP should be based on a assive measurement aroach. B. Parallel Data Transfer GridFTP can establish multile TCP connections in arallel by using OPTS ET or OPTS STO command. Hence, a single file can be transferred via multile TCP connections from/to a single or multile GridFTP servers. Higher throughut than with a single TCP connection can be exected through aggregation of multile TCP connections [6] [8]. This can be exlained by the following three reasons. First, larger bandwidth can be gained by aggregating multile TCP connections when cometing with other TCP connections in the TCP congestion avoidance hase [17]. This is because AIMD window flow control is adoted in the TCP congestion avoidance hase and data transfer can be erformed better by aggregating multile TCP connections in a network with a low acket loss robability. Second, the total TCP socket buffer size that can be used in a file transfer becomes large by aggregating multile TCP connections. This is because the total of TCP socket buffer sizes is times larger by aggregating TCP connections. Third, the start-u of the transfer rate is accelerated in the TCP slow start hase by aggregating multile TCP connections. In the slow start hase, the congestion window doubles for every round-tri time. Accordingly, the start-u for the transfer rate is times faster by aggregating TCP connections. However, if the number of aggregate TCP connections is too large, it results in decreased throughut for the following reasons. First, the window size er TCP connection decreases, and TCP timeout are more likely to occur. Second, the overhead required for the GridFTP server and client to rocess the TCP rotocol stack increases. Accordingly, the otimal value for the number of aggregate TCP connections,, must be determined according to several network conditions. However, it has not been fully investigated and is still an unresolved roblem how to determine the number of arallel TCP connections in various network environments. III. AALYTIC MODEL In this aer, modeling of GridFTP is erformed using a network modeling technique roosed in [11], [18]. In what follows, a rimary feature of GridFTP, arallel data transfer is modeled.
GridFTP server TCP model TCP model TCP model w data acket arallel data transfer (multile TCP connections) + + router ACK acket router model B GridFTP client ẇ =, (t) =, and w(t) = w(t ) = w, so the following relationshi is obtained from Eq. (1). 3 ( 3 + 2 w 2) + (3 2 w ) w T O 3 = (2) In the above equation, TCP window size w in steady state is assumed to be larger than or equal to 3. Under this assumtion, the robability of detecting acket loss due to TCP timeouts, T O, is given by [19] T O 3 w (3) flow distribution Fig. 1. Modeling GridFTP with arallel data transfer by aggregating TCP continuous-time models By solving in Eqs. (2) and (3) for w, TCP window size w in steady state is obtained as w 1 ( ) 6 + 21 3 + (4) 2 In our, a GridFTP server sends a file to a GridFTP client using arallel TCP connections. Other cases of GridFTP client-to-server file transfers can be easily modeled using the same modeling aroach. GridFTP suorts arallel data transfer, artial file transfer, and third-arty control of data transfer, so there exists one or more GridFTP servers for a single GridFTP client. In our, traffic on the control channel is assumed to be negligible, and only traffic on data channel is modeled. Modeling GridFTP is erformed as follows (Fig. 1). First, the GridFTP server is modeled by aggregating multile continuous-time models of the TCP congestion control mechanism. When GridFTP erforms arallel data transfer, multile TCP connections are established between the GridFTP server and client. Accordingly, multile TCP congestion control mechanism models are aggregated at the GridFTP server. IV. STEADY STATE AALYSIS In this section, the otimal number of arallel TCP connections is derived by erforming steady state for our GridFTP model. First, a case when the TCP socket buffer size W is larger than the bandwidth-delay roduct er TCP connection (i.e., TCP throughut round-tri time) is considered. In this case, according to [11], [18], changes in the TCP window size w(t) at time t including a TCP timeout mechanism are modeled using a fluid-flow aroximation as w(t ) ẇ = (1 (t)) w(t) (t) 2 ) w(t)w(t {1 T O (t)} 3{ } 4 w(t ) (t) 3 w(t) 1 T O (t) (1) where (t) is the acket loss robability in a network at time t, is the round-tri time of TCP connections, and T O (t) is the robability of detecting acket loss at time t due to TCP timeouts. The acket loss robability and TCP window size in steady state are denoted by and w, resectively. In steady state, TCP throughut T in steady state is then given by T w 1 ( ) 6 + 21 3 + 2 Since TCP socket buffer size is assumed to be larger than the bandwidth-delay roduct er TCP connection, acket loss is exected to occur only because of network congestion. Let be the number of TCP connections, and B the bottleneck link bandwidth. Under these conditions, total throughut for all TCP connections T equals the bottleneck link bandwidth B in steady state. Therefore, B = T 2 ( 3 + ) 6 + 21 is obtained from Eq. (5). By solving this equation for, the acket loss robability in steady state is derived as ( 2 + 2 B + 2 ( ) ) 2 1 B (7) 3 From the above equation, one can find that the acket loss robability in a network in steady state is determined only by the bandwidth-delay roduct er TCP connection (i.e., B /). When the acket loss robability in steady state is and the throughut for all TCP connections is T, only a fraction of ackets are discarded in the network. So, effective throughut (i.e., goodut) for a TCP connection G is given by (5) (6) G T (1 ) (8) ext, a case when the TCP socket buffer size W is smaller than the bandwidth-delay roduct er TCP connection (TCP throughut round-tri time). In this case, TCP window size cannot be fully increased regardless of the available bandwidth. Accordingly, TCP throughut T is limited by T = W When TCP socket buffer size W is smaller than the bandwidth-delay roduct er TCP connection, the bottleneck link bandwidth cannot be utilized at 1%. Provided that the (9)
acket loss caused by background traffic is negligible, the acket loss robability in steady state is given by = (1) Thus, the effective throughut for a TCP connection G in steady state becomes G T (1 ) = T = W (11) From Eqs. (8) and (11), the GridFTP goodut (total effective throughut for all TCP connections) G is given as a function of the TCP socket buffer size W, the number of arallel TCP connections, round-tri time, and the bottleneck link bandwidth B; i.e., ( W G min, (1 ( )) ) 6 + 21 3 + (12) 2 Thus, the otimal number of arallel TCP connections is obtained by determining that maximizes Eq. (12); i.e., from Eq. (12), the otimal value of is derived as ( = 3 B 3 W 3 ) 9 B 2 2 16 B W + 7 W 2 B (13) 9 B 6 W On the contrary, Eq. (12) indicates that for a given number of arallel TCP connections,, the TCP socket buffer size W should be large enough to maximize G in Eq. (12). In ractice, the TCP socket buffer size W should not be too large for reventing unnecessary memory consumtion. Hence, the TCP socket buffer size W should be as large as ossible, but no more than the bandwidth-delay roduct of the network. V. UMEICAL EXAMPLES In what follows, the effect of the number of arallel TCP connections and TCP socket buffer size on GridFTP erformance has been quantitatively investigated through several numerical examles of the steady state. First, GridFTP goodut (total goodut of all TCP connections) G in steady state is shown in Fig. 2. In this figure, we use the following arameters: the bottleneck link bandwidth B = 8.3 [acket/ms] (corresonding to 1 [Mbit/s] when a acket size is 15 [byte]) and TCP socket buffer size W = 64 [Kbyte]. The round-tri time for TCP connections and the number of arallel TCP connections are varied. In addition, the acket loss robability for GridFTP in steady state is lotted in Fig. 3. The followings regarding GridFTP goodut can be observed from these figures. First, from Fig. 2, the number of arallel TCP connections should be large for utilizing the bottleneck link bandwidth at almost 1%. The required number of arallel TCP connections for full link utilizing is almost roortional to the round-tri time for TCP connections. This henomenon is in agreement with results (see, e.g., [8]). Second, GridFTP goodut decreases slightly with the further increase in the number of arallel TCP connections. This tendency aears obviously for a small round-tri time. G 8 6 4 2 5 1 15 1 3 5 4 Fig. 2. GridFTP goodut G (effect of round-tri time and the number of arallel TCP connections ) (B = 8.3 [acket/ms], W = 64 [Kbyte]).15.1.5 5 1 15 1 3 5 4 Fig. 3. GridFTP acket loss robability (effect of round-tri time and the number of arallel TCP connections ) (B = 8.3 [acket/ms], W = 64 [Kbyte]) This is because the acket loss robability increases as the number of arallel TCP connections increases and/or roundtri time for TCP connections decreases [8], as can be seen from Fig. 3. We then focus on the effect of the number of arallel TCP connections and TCP socket buffer size W on GridFTP goodut and the acket loss robability in steady state. The GridFTP goodut and acket loss robability for different numbers of arallel TCP connections and TCP socket buffer sizes W is shown in Figs. 4 and 5, resectively. The bottleneck link bandwidth is set to B = 8.3 [acket/ms], and the roundtri time for TCP connections at = 1 [ms]. From Fig. 4, one can find that the TCP socket buffer size W and the number of arallel TCP connections must be
1.75 G.5.25 W 4 6 1 3 5 4 G 15 1 5 5 1 15 1 3 5 4 Fig. 4. GridFTP goodut G (effect of the number of arallel TCP connections and TCP socket buffer size W ) (B = 8.3 [acket/ms], = 1 [ms]) Fig. 6. GridFTP goodut ( G ) (effect of round-tri time and the number of arallel TCP connections ) (B = 16.6 [acket/ms], W = 64 [Kbyte]).4.3.2.1 W 4 6 1 3 5 4 Fig. 5. GridFTP acket loss robability (effect of the number of arallel TCP connections and TCP socket buffer size W ) (B = 8.3 [acket/ms], = 1 [ms]) aroriately configured for fully utilizing the bottleneck link bandwidth. For instance, when the TCP socket buffer size W is small (e.g., 16 [Kbyte]), the number of arallel TCP connections should be very large. Accordingly, configuring the TCP socket buffer size W to a sufficiently large value is desired for avoiding an extremely large number of arallel TCP connections. As the number of arallel TCP connections increases, the acket loss robability for GridFTP increases according to Fig. 5. Also, the acket loss robability for GridFTP is indeendent of the TCP socket buffer size W. Based on these observations, a guideline for tuning GridFTP control arameter is to allocate as large a TCP socket buffer size W as ossible, and to establish the number of arallel TCP connections that can fully utilize the bottleneck link bandwidth. ote that the bandwidth-delay roduct of the network is in ractice the uer-limit of the TCP socket buffer size W. ext, effect of the bottleneck link bandwidth on the otimal arameter configuration of GridFTP is investigated. Figure 6 is a result with a larger bottleneck link bandwidth B (B = 16.6 [acket/ms]) than that of Fig. 2. From this figure, it can be found that the number of arallel TCP connections should be increased accordingly when the bottleneck link bandwidth is increased. However, by comaring Figs. 2 and 6, one can find that goodut degradation for GridFTP for a large number of arallel TCP connections is smaller for a larger bottleneck link bandwidth. This is because the bandwidth-delay roduct of each TCP connection increases as the bottleneck link bandwidth increases, and, consequently, TCP timeouts less likely to occur. This means that arameter configuration for GridFTP is simler in a faster or wider-area network. Finally, the validity of our aroximation is examined through comarison of and results. ns-2 simulator (version 2.28) [] is used for all s. A simle network toology of one ho is used in the. The bottleneck link bandwidth is set to B = 8.3 [acket/ms], and the two-way roagation delay at τ = 1 [ms] or τ = [ms]. The acket size is fixed at 1,5 [byte], and GridFTP server is modeled by aggregating FTP traffic sources. Every is run for 6 [s] while changing the number of arallel TCP connections (i.e., the number of active FTP traffic sources), and the goodut and acket loss robability are measured. First, results for the two-way roagation delay τ = 1 [ms] of the bottleneck link are shown in Figs 7 and 8. These figures show the goodut and acket loss robability for GridFTP when the number of arallel TCP connections is changed. Also, analytic results, which are calculated based
GridFTP Goodut: G * [Mbit/s] 1 1 8 6 4 1 3 4 5 umber of TCP Flows: GridFTP Goodut: G * [Mbit/s] 1 1 8 6 4 1 3 4 5 umber of TCP Flows: Fig. 7. umber of arallel TCP connections vs. GridFTP goodut (B = 8.3 [acket/ms], τ = 1 [ms], W = 64 [Kbyte]) Fig. 9. umber of arallel TCP connections vs. GridFTP goodut (B = 8.3 [acket/ms], τ = [ms], W = 64 [Kbyte]) Packet loss robability: *.2.15.1.5 Packet loss robability: *.1.8.6.4.2 1 3 4 5 umber of TCP connections: 1 3 4 5 umber of TCP connections: Fig. 8. umber of arallel TCP connections vs. GridFTP acket loss robability (B = 8.3 [acket/ms], τ = 1 [ms], W = 64 [Kbyte]) Fig. 1. umber of arallel TCP connections vs. GridFTP acket loss robability (B = 8.3 [acket/ms], τ = [ms], W = 64 [Kbyte]) on the average round-tri time for TCP connections obtained by, are lotted. From these figures, one can find that the GridFTP goodut and acket loss robability are accurately estimated by our steady state. Simulation results for a small two-way roagation delay of the bottleneck link (τ = [ms]) are shown in Figs. 9 and 1. Comared to the case with larger two-way roagation delay (Figs. 7 and 8), it can be found that analytic results deviate from results, in articular, when there are a large number of arallel TCP connections (e.g., = 5). This henomenon can be exlained by the following reason. In a network with a small bandwidth-delay roduct and a large number of arallel TCP connections, window size for each TCP connection becomes small so that TCP timeouts are more likely to occur. However, the robability of detecting acket loss due to TCP timeouts, T O, is aroximated by Eq. (3), which should not be used when the acket loss robability is large [19]. More accurate modeling of the robability of detecting acket loss due to TCP timeouts, T O, is necessary, but it is beyond the scoe of the current aer. However, it should be noted that our steady state is sufficiently usable in ractice for otimizing control arameter for GridFTP. amely, recall that the guideline for GridFTP arameter tuning is to use a sufficiently large TCP socket buffer size W and to configure the number of arallel TCP connections for fully utilizing the bottleneck link bandwidth. Hence, effect of the modeling error in an extremely larger number of arallel TCP connections can be negligible for arameter configuration uroses. VI. COCLUSIOS AD FUTUE TOPICS In this aer, we have investigated the otimal arameter configuration for GridFTP, i.e., the number of arallel TCP connections and TCP socket buffer size, by erforming a steady state for GridFTP. First, a continuous-time model for GridFTP has been derived by aggregating multile TCP continuous-time models. Steady state erformance metrics (e.g., GridFTP goodut and acket loss robability) have been derived using our GridFTP model. By focusing on
the number of TCP connections and TCP socket buffer size, we have also derived the otimal arameter configuration for GridFTP, and quantitatively show erformance limitations of GridFTP through analytic and results. Our GridFTP arameter tuning guideline is to allocate as large a TCP socket buffer size W as ossible (but no more than the bandwidthdelay roduct of the network) and to establish the number of arallel TCP connections that can fully utilize the bottleneck link bandwidth according to Eq. (13). We have also validated our aroximate by comaring results with analytic ones. Future research toics include imroving the accuracy of our aroximate (e.g., accurate modeling of TCP timeout mechanism) and erforming of GridFTP in more generic network configurations with, for instance, lossy link, background traffic, and heterogeneous TCP connections. Also imortant is alying our analytic results for automatically otimizing GridFTP erformance. Our work on designing an automatic arameter configuration mechanism for GridFTP, which utilizes our analytic results and has comatibility with existing GridFTP servers, will be ublished soon [21], [22]. [18] S. H. Low, F. Paganini, and J. C. Doyle, Internet congestion control, IEEE Control Systems Magazine, vol. 22, no. 1,. 28 43, Feb. 2. [19] J. Padhye, V. Firoiu, D. Towsley, and J. Kurose, Modeling TCP throughut: a simle model and its emirical validation, in Proceedings of ACM SIGCOMM 98, Set. 1998,. 33 314. [] The network simulator ns2, available at htt://www.isi.edu/nsnam/ ns/. [21] T. Ito, H. Ohsaki, and M. Imase, Automatic arameter configuration mechanism for data transfer rotocol GridFTP, in Proceedings of the 6 International Symosium on Alications and the Internet (SAIT 6), Jan. 6,. 32 38. [22], GridFTP-APT: Automatic arallelism tuning mechanism for data transfer rotocol GridFTP, in Proceedings of 6th IEEE International Symosium on Cluster Comuting and the Grid (CCGrid6), May 6,. 454 461. EFEECES [1] J. Postel, Transmission control rotocol, equest for Comments (FC) 793, Set. 1981. [2] W. Allcock et al., GridFTP: Protocol extensions to FTP for the Grid, OGF Document Series GFD., Ar. 3, also available as htt://www. ggf.org/documents/gfd..df. [3] J. Semke, J. Mahdavi, and M. Mathis, Automatic TCP buffer tuning, in Proceedings of ACM SIGCOMM 98, vol. 28, Oct. 1998. [4] T. Dunigan, M. Mathis, and B. Tierney, A TCP tuning daemon, in Proceedings of SuerComuting: High-Performance etworking and Comuting, ov. 2. [5] S. Thulasidasan, W. Feng, and M. K. Gardner, Otimizing GridFTP through dynamic right-sizing, in Proceedings of IEEE International Symosium on High Performance Distributed Comuting, June 3,. 14 23. [6] H. Sivakumar, S. Bailey, and. L. Grossman, PSockets: The case for alication-level network striing for data intensive alications using high seed wide area networks, in Proceedings of the ACM/IEEE Conference on Suercomuting, ov.. [7] T. J. Hacker, B. D. Athey, and B. oble, The end-to-end erformance effects of arallel TCP sockets on a lossy wide-area network, in Proceedings of the 16th IEEE-CS/ACM International Parallel and Distributed Processing Symosium (IPDPS), Ar. 2,. 434 443. [8] L. Qiu, Y. Zhang, and S. Keshav, On individual and aggregate TCP erformance, in Proceedings of Internetl Conference on etwork Protocols, Oct. 1999,. 3 212. [9] D. Lu, Y. Quao, P. Dinda, and F. Bustamante, Modeling and taming arallel TCP on the wide area network, in Proceedings of the 19th IEEE International Parallel and Distributed Processing Symosium, Ar. 5. [1]. Cohen and S. amanathan, TCP for high erformance in hybrid fiber coaxial broad-band access networks, IEEE/ACM Transactions on etworking, vol. 6,. 15 29, Feb. 1998. [11] H. Ohsaki, J. Ujiie, and M. Imase, On scalable modeling of TCP congestion control mechanism for large-scale IP networks, in Proceedings of IEEE SAIT 5, Feb. 5,. 361 369. [12] Global Grid Forum, htt://www.ggf.org/. [13] J. Postel and J. eynolds, File transfer rotocol (FTP), equest for Comments (FC) 959, Oct. 1985. [14]. Elz and P. Hethmon, FTP security extensions, equest for Comments (FC) 2228, Oct. 1997. [15] P. Hethmon and. Elz, Feature negotiation mechanism for the file transfer rotocol, equest for Comments (FC) 2389, Aug. 1998. [16] Globus Toolkit, available at htt://www.globus.org/. [17] S. Floyd, Highseed TCP for large congestion windows, Internet Draft draft-ietf-tsvwg-highseed-1.txt, Aug. 3.