On 25 Sep 2000, Henner Eisen wrote:
> >>>>> "Donald" == Donald Becker <becker@xxxxxxxxx> writes:
>
> Donald> This is one of the reasons for drivers to keep their Tx
> Donald> queues to a reasonable length. Two years ago an
> Donald> almost-always-sufficient Tx queue length was between 6 and
> Donald> 10, so most of my bus-master drivers use a queue length of
> Donald> 10. Ten 1500 byte packets at 10Mbps is still not too
> Donald> long, but ten 60 byte packets (approx. 96 byte periods on
> Donald> the wire) isn't very long at 100Mbps.
>
> Would it make sense to use the data size (the accumulated
> skb->len of the sk_buffs in a packet schedulerīs queue) instead of the
> current tx queue length (number of sk_buffs) for mesuring the packet
> schedulerīs queues?
The data size frequently isn't a better metric:
- On a busy half duplex network your outbound queue delay is a primarily
function of the number of packets, not the size of those packets.
- Is it really better for your "high priority" packet to be waiting behind
50 tinygrams rather than only 4 large packets?
There might be a "have times changed?" issue here:
Are most high performance networks switched? Probably yes.
Does switching (full duplex) change the queue metric? Perhaps..
Will Ethernet flow control change this back? Probably yes.
Donald Becker becker@xxxxxxxxx
Scyld Computing Corporation http://www.scyld.com
410 Severn Ave. Suite 210 Beowulf-II Cluster Distribution
Annapolis MD 21403
|