1.3 GESTIÓN ADMINISTRATIVA DE CREDI HOGAR
1.3.1 Aspectos Constitutivos
Despite preservation of data locality, the efficiency of remote data transfers in WANs is vital to the perfor- mance of data-intensive applications, since they are inevitable in geo-distributed environments. Importantly, TCP/IP is still the dominant Internet protocol suite enabling remote data transfers. However, since initially designed for LANs, TCP exhibits its shortcomings in WANs which cause serious performance deterioration. Lakshman and Madhow (1997) recognize that the classic TCP design (Jacobson, 1988) discriminates against
connections with high latency and overreacts to random packet loss. This behavior is unfriendly to remote data transfers as they typically show high latency and the tendency of packet loss. Further, the authors reveal that large product of bandwidth and network delay on a TCP connection, referred to as bandwidth-delay product (BDP), may overflow buffers of network devices along the network path and further deter TCP performance with increasing packet loss. Mathis et al. (1997) manifest the problem by formulating a mathematical model to describe the TCP fast retransmit mechanism as follows:
BW = M SS·C
RT T ·√p (2.1)
In this equation, BW, RT T, andpdenote bandwidth, round trip time (RTT), and the probability of packet loss, respectively. M SS andC are constant values, representing maximum segment size (MSS) and constant of proportionality. As reflected in the equation, the probability of packet loss is inversely correlated with bandwidth and RTT. In other words, high BDP will amplify packet loss, which in turn hurts throughput and degrades TCP performance. Padhye et al. (1998) refine the model of Mathis et al. (1997) by incorporating the TCP timeout mechanism. Floyd et al. (2000) propose the equation-based congestion control as a replacement for the best-effort, additive-increase/multiplicative-decrease (AIMD) mechanism using such mathematical models, which adjusts the sending rate as a function of measured rate of packet loss.
Centering on the macroscopic behavior of TCP congestion control, an array of TCP variants are devel- oped to cope with performance degradation observed in long-distance data transfers in WANs. The TCP Westwood (Mascolo et al., 2001) focuses on TCP congestion control for wireless networks with lossy links. It introduces a fast recovery mechanism that readjusts the congestion window consistently with the sender-side bandwidth measurement, in contrast with blindly halving the congestion window as performed by the TCP Reno, upon packet loss. In effect, it reduces the false positives of congestive packet loss and therefore alleviates the disturbance of packet loss to TCP throughput. The TCP Veno (Fu and Liew, 2003) adopts a similar approach to distinguish between congestive and random packet loss in wireless networks. Leith and Shorten (2004) devise the H-TCP for high-speed, homogeneous networks – they observe that the additive increasing factor in a high-speed, homogeneous network diverges from that in a conventional network, and they dynamically adjust the factor to accommodate different network settings and fully utilize available bandwidth. Xu et al. (2004) and Ha et al. (2008) propose to accelerate the additive increasing phase to overcome the lag in congestion window recovery caused by high network delay.
The prevailing approaches share a common assumption that packet loss is an indication of network congestion. The rationale behind this is that network congestion causes bandwidth oversubscription, and certain packets are forcibly dropped due to buffer overflow on intermediate network devices. However, as network capacity exponentially grows, the relationship between packet loss and network congestion becomes tenuous as identified by Cardwell et al. (2016). Hence, Cardwell et al. (2016) build a non-loss-based congestion avoidance model based on bandwidth and RTT and devise a new distributed congestion control mechanism named BBR which achieves optimal bandwidth utilization as a result.
An alternative to tackling TCP performance degradation caused by large BDP is to use multi-path TCP (MPTCP) initially proposed by Ford et al. (2013). The core idea is to perform a data transfer using parallel TCP sub-flows simultaneously to alleviate the negative impact of high latency and packet loss – congestion window growth is multiplied by parallel TCP sub-flows, while slowdown caused by packet loss is diluted among the multiple TCP sub-flows. This approach is commonly adopted by data transfer tools for long-distance data transfers. Raiciu et al. (2012) provide an in-depth explanation of the MPTCP implementation.
CHAPTER 3: Advancing Data-Driven Scientific Collaboration in Geo-Distributed Environment
In this chapter1, we present our work in the RADII project that facilitates data-driven scientific collabo- rations among geo-distributed research institutions. These scientific collaborations are common, especially in multidisciplinary projects, in which domain scientists are committed to specific areas but put forward the project collaboratively. Typically, the collaborations are realized in the form of data-intensive applications such as data processing workflows and pipelines consisting of computational jobs running in a geo-distributed fashion among the participating institutions, which desperately call for streamlined infrastructure support in the geo-distributed environment.
The advance in national CI and cloud computing has dramatically improved the accessibility of physical infrastructure in the geo-distributed environment. However, domain scientists still face the prevalent chal- lenges in geo-distributed systems when deploying and executing data-intensive applications for scientific collaborations. Mainly, they lack the high-level abstraction and mechanism for describing, constructing, and managing such collaborations with simplicity and accuracy. Furthermore, data sharing among geo-distributed locations tends to create the significant performance bottleneck due to the absence of network abstraction and optimization.
Hence, we have developed RADII as an integrated solution to address these challenges. We have introduced a simple yet powerful data model for describing scientific collaborations. We have also built the software infrastructure that maps the high-level collaboration description to low-level, virtualized infras- tructure primitives for deploying and orchestrating the collaboration on the physical infrastructure at a fine granularity. In particular, we use software-defined networking (SDN) to abstract the network. On top of that, we have devised an MCF-based optimization approach and developed a TCP enhancement to improve the bandwidth utilization and accelerate remote data transfers in the collaboration.
1
Content of this chapter previous appeared in preliminary form in the following paper:
Jiang, F., Castillo, C., and Schmitt, C. (2016). RADII: Bridging the Divide Between Data and Infrastructure Management to Support Data-Driven Collaborations. In Big Data (Big Data), 2016 IEEE International Conference on, pages 370–377. IEEE.
For the rest of this chapter, we first introduce the high-level data model (Section 3.1) and the software infrastructure (Section 3.2) that support the scientific collaborations. Then, we dive into the SDN-based network abstraction (Section 3.3) and explain the MCF-based network optimization (Section 3.4) and TCP enhancement (Section 3.5) in detail. Lastly, we present the experimental evaluation (Section 3.6) and summarize this chapter (Section 3.7).