Excess buffer requirement for EPD schemes in ATM networks

It is known that the performance of TCP over ATM can be significantly degraded if the bandwidth is occupied by cells belonging to packets that are already corrupted by cell loss due to buffer overflow. The Early Packet Discard (EPD) mechanism is a well known technique to improve the performance of TCP or other transport layer protocols over ATM by allowing only non-corrupted packets to be transmitted. The basic idea of EPD is to discard an entire packet prior to buffer overflows; a switch will drop incoming packets whenever the buffer occupancy exceeds a pre-defined EPD threshold. The excess buffer capacity for the EPD mechanism refers to the available buffer capacity above the EPD threshold in a switch. The excess buffer is needed to accommodate the outstanding cells that are not discarded. Moreover, sufficient excess buffer is important in maintaining good throughput performance. In this article, we shall present analytical results on the worst-case excess buffer capacity for EPD mechanism that incorporate per-VC accounting and per-VC queueing techniques, extending the work of Turner (J.-S. Turner, Maintaining High Throughput During Overloading In ATM Switches, IEEEINFOCOM'96, March 1996) for the plain EPD case. We derive tight upper bounds on the excess buffer occupancy in per-VC based EPD schemes. Our analytical results show that per-VC based EPD schemes can provide better performance at the expense of more excess buffer than the plain EPD scheme. Simulation results are also presented and shown to be in agreement with the theoretical results. We also provide a simulation study on more general cases for both plain EPD and per-VC based EPD schemes, and reasonable rules of thumb for choosing excess buffer in these situations are discussed.