As network technologies continue to improve, the bottleneck for end-to-end data transfer is expected to shift from the network to the end-system. We demonstrate this bottleneck exists by transferring data over a 1 Gbps dedicated channel between two end-systems. Our experiments show that the average throughput is substantially lower than the network transmission line rate when the receiving end-system load is 1 or more (in other words, it performs tasks in addition to network I/O.) This is mainly because transport protocols are unaware of the end-system bottleneck, which leads to a sawtooth pattern of increasing and throttling of the sending rate. We then introduce a mechanism to estimate the best rate at which an end-system may perform network I/O, which we call the effective end-system bottleneck rate. We show that the end-system bottleneck not only depends on its hardware specifications, but also the workload it is currently executing. Our model estimates the effective end-system bottleneck rate by performing the following steps: (i) determining the expected service time of the receiving process in the midst of other competing processes,(ii) applying the above service rate for packet loss analysis for NIC buffer, the RX ring buffer and the socket buffer, and (iii) determining the sending rate which yields the minimum transfer time for a file on a specified path, given the above packet loss.
Introspective End-system Modeling to Optimize theTransfer Time of Rate Based Protocols
SERAZZI, GIUSEPPE
2011-01-01
Abstract
As network technologies continue to improve, the bottleneck for end-to-end data transfer is expected to shift from the network to the end-system. We demonstrate this bottleneck exists by transferring data over a 1 Gbps dedicated channel between two end-systems. Our experiments show that the average throughput is substantially lower than the network transmission line rate when the receiving end-system load is 1 or more (in other words, it performs tasks in addition to network I/O.) This is mainly because transport protocols are unaware of the end-system bottleneck, which leads to a sawtooth pattern of increasing and throttling of the sending rate. We then introduce a mechanism to estimate the best rate at which an end-system may perform network I/O, which we call the effective end-system bottleneck rate. We show that the end-system bottleneck not only depends on its hardware specifications, but also the workload it is currently executing. Our model estimates the effective end-system bottleneck rate by performing the following steps: (i) determining the expected service time of the receiving process in the midst of other competing processes,(ii) applying the above service rate for packet loss analysis for NIC buffer, the RX ring buffer and the socket buffer, and (iii) determining the sending rate which yields the minimum transfer time for a file on a specified path, given the above packet loss.I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
