We reconsider several issues related to fragmentation and reassembly in IP. W e rst reconsider reassembly. W e describe a simple expected case optimization that improves reassembly performance to 38 instructions per fragment if the fragments arrive i n F I F O order (the same assumption made in header prediction) which has been implemented in the NetBSD kernel. Next, we i n troduce the new idea of Graceful Intermediate Reassembly (GIR), w h i c h i s a g e neralization of the existing IP

more »

... sms of destination and hop-by-hop reassembly. In GIR, we coalesce the fragments at an intermediate router in order to use the largest sized packets on its outgoing interface. We s h o w that GIR always outperforms hop-by-hop reassembly and can be implemented economically in routers with small processing and memory costs. We then reconsider fragmentation. We s h o w t h a t a voiding fragmentation has costs of its own in terms of increased packet processing and/or round-trip delays. We describe measurements in which TCP performance improves after turning on fragmentation. For example, on Ethernet under NetBSD using a 536 byte segment size for TCP we get a throughput of only about 5.5Mb/s, whereas we get a throughput a of 8.45Mb/s using a segment size of 1460 bytes (without fragmentation) and a throughput of 8.82Mb/s using a segment size of 16260 bytes with fragmentation. We also describe a simple performance model that can be used to determine when fragmentation is bene cial. Finally, w e address the major disadvantages of fragmentation. A major problem is that when a fragment is lost, the entire TCP segment must be retransmitted, resulting in reduced or zero goodput under loss. We describe a mechanism | dynamic segment sizing, in which the segment size is dynamically reduced after loss | which addresses this problem. We describe simulations (using our modi ed NetBSD kernel) which shows that dynamic segment s i z i n g k eeps the goodput at reasonable levels even under extremely lossy conditions. All our mechanisms (reassembly optimization, GIR, and dynamic segment sizing) are orthogonal and can be applied to other protocol suites besides TCP/IP. a 1 Mb/s = 1048576 bits/second