| My guess is that a few of those routers are congested and/or using a switch traffic flow cache where entries get replaced in milliseconds. When this happens, as with the first packet sent to a destination, the hardware switch fabric can't quickly move the packet, it has to be examined by the CPU. The routing table is used to find the best egress interface, the packet is sent there, and the switch cache gets updated with this destination. Probably some other destination got kicked out of the cache. The problem is that the CPU gets backed up with work and its queue of packets to handle the slow way overflows. The queue length can be tuned, and perhaps its not the best length. Its far better to drop data packets than routing packets, and they may hold space for them. Anyway, this is probably why medium speed transmission gets drops and higher speed not.
There could be a buffer carving issue on the router - lots of small buffers and large buffers and not enough medium sized ones. On the Internet, the greatest number of packets are 64 bytes long (ACKs, ICMP, SYNs etc), next would be full Ethernet frame sized packets. The medium sized packets are not nearly so common, and so one or more routers may be short on buffers that size.
There's not much you can do, however. Hope the backbone network guys notice all the drops and feel like tweaking router parameters after getting approval to do it during a future maintenance window! Not going to happen very often. |
