FIBRE CHANNEL and IP SAN INTEGRATION(12)

The maturity and mission-critical deployment of Fibre Channel (FC) in storage area networks (SANs) creates a unique class of multi-terabit networks with demanding throughput, latency, scalability, robustness, and availability requirements. This paper revie

respect to delay and error recovery is an area that needs further research and understanding.4.4 Long Fast Network ExperimentIn another case [30], the University of Tokyo conducted an experiment using ‘iperf’ running TCP between Maryland and Tokyo, traversing the Abilene and APAN networks. The result was surprising in that Fast Ethernet is sometimes faster than Gigabit Ethernet on LFN. The main cause of the throughput degradation with Gigabit Ethernet LFN tests was congestion overflow of an intermediate router, resulting in cranking of TCP time out, slow start and congestion control mechanisms. Transmission rate control is important to mitigate the overflow in the bottleneck’s buffer in addition to the window size control. Therefore, transmit rate or bandwidth limiting is another important mechanism that avoids or mitigates the impact of congestion overflow in an intermediate network.4.5 Fast WriteWe examine a method to improve the write performance over a long delay network. As shown in Table 1, a SCSI write transaction incurs two round trip delays for a data block. The maximum block size is determined by the target (storage) device’s buffer capacity and is specified by the target in the XFR_RDY message. For example, writing one MB of data using 64 KB blocks takes 16 transactions, which is 32 round trips plus data transfer time. Fast Write [31] is a way to minimize round-trip delay overhead and accelerate SCSI write performance leveraging a gateway’s buffer capacity. The XFR_RDY is spoofed by the gateway on the initiator (server) side of the network, and the data is buffered by the gateway on the target side of the network until the target sends its own XFR_RDY. In addition, the use of TCP protocol with selective retransmission (on error) provides better frame loss recovery than retransmitting the entire block on timeout (as in the SCSI-FCP case). With Fast Write, the number of round trip involved for a 1 MB transfer is reduced to two round trips.Figure 8 Fast Write Example112