Special Report: Network Provisioning

In meeting the DOE large-science requirements, it is instructive to note that the end users view the network as a tool or a resource much like a computer. Their main goal is to conduct scientific activities in their areas with minimal demands for using the network. Over the past years, however, several users have become (not always willingly) network experts in their attempts to scale TCP or other protocols to the required throughput levels using methods such as parallel streams and buffer tuning. But such improvements require in-depth knowledge of the protocols and are achieved by significant efforts by groups of experts. Such a "wizard gap" exists at all levels and the expertise needed for such efforts is beyond a typical science user. In a nutshell, the "gray matter tax" for such efforts is undesirably high. Thus, one of the considerations in meeting the DOE large-science needs is to advance the state of network protocols to make them plug-and-play for the application users. In particular, the use of protocols must be transparent to the users; they just specify the performance requirements for connections and all the other details such as the underlying provisioning or the corresponding parameter values must be hidden from them. It is to be noted that achieving such a level of transparency in the presence of a multitude of provisioning modes and matching protocols represents a significant challenge.

5.1 Existing Transport Protocols The limitations of current transport methods, particularly TCP, in addressing high performance transport applications, have prompted a number of solutions with varying degrees of successes. To a large extent these efforts are focused on IP-based protocols such as the various TCP enhancements, net100, HSTCP, STCP, FAST, and UDP-based methods such as tsunami and SABUL. There are also efforts to adapt the protocols designed for Storage Area Networks (SAN), such as Fiber Channel, to the wide-area networks. Protocols that are specifically optimized to exploit the properties of dedicated channels are quite limited. Since a major consideration of IP- based protocols is the impact on other traffic streams, a significant effort has been extended to ensure their "gracefulness" or "fairness". Lack of this consideration in dedicated channels opens up a vast potential for customization and optimization of the transport protocols, if not, motivating a whole new approach to their design.

UDP burst Figure 7. Response of TCP variants to a UDP burst. The problem of optimizing TCP to achieve ultra high throughputs is extremely complicated due the high non-linearity of its dynamics. By suitably controlling the slow-start phase and AIMD parameters, it is quite possible overcome some TCP limitations. But optimizing certain measures of performance might result in degrading the others; for example, as shown in Figure 7 avoidance of overshoot during slow start might result in slower recovery. Among the TCP based methods presently under investigation are various versions of TCP (Reno, Vegas), HSTCP, STCP, XCP, net100, and FAST. Protocols that optimize the flow rates of UDP streams achieve high utilization of the connection bandwidths, for example, tsunami, SABUL, IQ-RUDP, hurricane, RBUDP, FOBS, and IUDP, but they often significantly degrade the performance of competing TCP traffic. The class of protocols that adapt methods used in SANs, such as STP and Fiber Channel, typically achieve much higher data rates since the distance involved are much smaller and there is limited or no competing traffic. Recent efforts in Fiber Channel over SONET focus on utilizing SONET links as carriers for Fiber Channel streams. In addition to the protocol designs, their actual implementations have a large impact on the performance. Since link bandwidths are several Gbps or 10s of Gbps, and current off-the shelf NICs are typically operate at 1Gbps, striping methods are needed to utilize multiple NICs to generate aggregate throughputs commensurate with the link speeds. Also, these data rates are significantly higher than processor speeds, and hence methods are needed to minimize the impact on the processors by utilizing the OS-bypass methods such as RDMA.

Previous    Next    Table of Content for report: Network Provisioning    Home

Network Provisioning

 

 

 

Photuris.com - Optical Data Networking