Section: Scientific Foundations

End-to-end high performance and deterministic transport and service differentiation

Participants : Pascale Vicat-Blanc Primet, Jingdi Zeng, Dino Lopez Pacheco, Cong-Duc Pham.

In TCP/IP networks, the end-to-end principle aims at simplifying the network level while pushing all the complexity on the end host level. This principle has been proved to be very valuable in the context of the traditional low capacity Internet. In packet networking, congestion events are the natural counterpart of the flexibility to interconnect mismatched elements and freely multiplex flows. Managing congestion in packet networks is a very complex issue. This is especially true in IP networks where, at best, congestion information is very limited (e.g., ECN) or, at worst, non-existent, forcing the transmitter to infer it instead (e.g., based on losses or delay) in TCP.

The conservative behavior of TCP with respect to congestion in IP networks (RFC 2581) is at the heart of the current performance issues faced by the high-performance networking community. Several theoretical and experimental analysis have shown that the dynamics of the traditional feedback based approach is too low in very high speed networks that may lose packets. Consequently network resource utilization is not optimal and the application performance is poor and disappointing. Many Grid-enabled computing applications wish to transfer large volumes of data over wide area networks and require high data rates in order to do so. However, Grid-enabled applications are rarely able to take full advantage of the high-capacity (2.5 Gbit/s, 10 Gbit/s and upwards) networks installed today. Recent data for Internet 2 show that 90% of the bulk TCP flows (defined as transfers of at least 10 Megabyte of data) use less than 5 Mbit/s, and that 99% use less than 20 Mbit/s out of the possible 622 Mbit/s provision. There are many reasons for such poor performance. Many of the problems are directly related to the end system, to the processor and bus speed, and to the NIC with its associated driver. TCP configuration (e.g., small buffer space or features such as SACK being improperly negotiated) will have a significant impact. TCP itself was designed first and foremost to be robust and when congestion is detected, TCP accommodates the problem but at the expense of reduced performance. There are also design problems with TCP itself. For example, for a standard TCP connection with 1500-byte packets and a 100 ms round-trip time, achieving a steady-state throughput of 10 Gbit/s would require an average congestion window of 83,333 segments, and a packet drop rate of at most one congestion event every 5,000,000,000 packet (or equivalently, at most one congestion event every 1 2/3 hours). HighSpeed TCP   [47] and Scalable TCP   [54] increase the aggressiveness in high-throughput situations while staying fair to standard TCP flows in legacy contexts. FAST  [53] leverages the queueing information provided by round-trip time variations, in order to efficiently control buffering in routers and manage IP congestion optimally. These propositions are actively analyzed and experimented by the international community. RESO participates to the elaboration of a survey on protocols other than standard TCP in the framwork of the Data Transport research group of the Global Grid Forum [41] . RESO is organizing in 2005, the third edition of the leading international workshop in this domain (see http://www.ens-lyon.fr/LIP/RESO/pfldnet2005 ). Several issues have been already enlightened. Considering the traditional feedback loop will not scale with higher rate level under loss or congesting traffic conditions, it seems judicious to start examining alternative radical solutions.

On the other hand, flows crossing the IP networks are not equally sensitive to loss or delay variations. Since several years, research effort has been spent to solve the problem of the heterogeneous performance needs of the IP traffic. A class of solutions considers that the IP layer should provide more sophisticated services than the simple best-effort service to meet the application's quality of service requirements. Quality of service has been studied in IP networks in the context of multimedia applications  [46] . Various complementary solutions have to be integrated to carry end-to-end quality of service to grid applications to assure an efficient usage of the interconnected computing resources  [49] . Solution like Diffserv exhibits three types of limitations we are considering:

• the end-to-end performance that the DiffServ standardized services provide have not been largely studied in real networks;

• when experiment shows that end-to-end connection can benefit from advanced DiffServ QoS network functionalities, their usage by individual flows is not straightforward;

• the deployment of DiffServ architecture presents different scaling problems. Alternative approaches are proposed to solve this issue.

Finally, tools for measuring the end-to-end performance of a path between two hosts are very important for transport protocol and distributed application performance optimization. Bandwidth evaluation methods aim to provide a realistic view of the raw capacity but also of the dynamic behavior of the interconnection that may be very useful to evaluate the time for bulk data transfer. Existing methods differ according to the measurements strategies and the evaluated metric. These methods can be active or passive, intrusive or non-intrusive. Non-intrusive active approaches, based on packet train or on packet pair provide available bandwidth measurements and/or the total capacity measurements. None of the proposed tools, based on these methods, enable the evaluation of both metrics, while giving an overview of the link topology and characteristics.