Hello experts!
A colleque found a strange effect when he measured the throughput on
TCP connections with varying block size.
Between 2 i386 boxes running NetBSD (-current of mid-may) values
around 1 Mbytes/s are measured for the whole range between 0 and
2k blocksize, except around 256 bytes, where it drops to 20 kBytes/s.
Can anybody explain this?
The appended tcpdump output shows that there are 4k transferred in
short time, than it waits for ~1/5 s. (inetd's block size for discards is 4k)
With 2 Alphas running OSF/1 the throughput is more or less constant.
The performance drop happens too if NetBSD is sender and OSF/1 receiver,
but not in the opposite direction.

best regards
Matthias

19:15:03.493097 134.94.206.43.1027 > zel024-z.discard: P 1:1201(1200)
ack 0 win 17520 <nop,nop,timestamp 16889 15918>
19:15:03.493249 zel024-z.discard > 134.94.206.43.1027: . ack 1201 win
17520 <nop,nop,timestamp 16940 16889>
19:15:03.494564 134.94.206.43.1027 > zel024-z.discard: . 1201:2649(1448)
ack 0 win 17520 <nop,nop,timestamp 16889 15918>
19:15:03.496022 134.94.206.43.1027 > zel024-z.discard: . 2649:4097(1448)
ack 0 win 17520 <nop,nop,timestamp 16889 15918>
19:15:03.690706 zel024-z.discard > 134.94.206.43.1027: . ack 4097 win
17520 <nop,nop,timestamp 16941 16889>
19:15:03.693065 134.94.206.43.1027 > zel024-z.discard: P 4097:5297(1200)
ack 0 win 17520 <nop,nop,timestamp 16890 15918>
19:15:03.693165 zel024-z.discard > 134.94.206.43.1027: . ack 5297 win
17520 <nop,nop,timestamp 16941 16890>
19:15:03.694523 134.94.206.43.1027 > zel024-z.discard: . 5297:6745(1448)
ack 0 win 17520 <nop,nop,timestamp 16890 15918>
19:15:03.695997 134.94.206.43.1027 > zel024-z.discard: . 6745:8193(1448)
ack 0 win 17520 <nop,nop,timestamp 16890 15918>
19:15:03.890709 zel024-z.discard > 134.94.206.43.1027: . ack 8193 win
17520 <nop,nop,timestamp 16941 16890>
19:15:03.893069 134.94.206.43.1027 > zel024-z.discard: P 8193:9393(1200)
ack 0 win 17520 <nop,nop,timestamp 16890 15918>