[netdrivers] latest on gige cards and performance?
Wed Aug 27 19:42:08 2003
On Wed, 27 Aug 2003, Donald Becker wrote:
> On Wed, 27 Aug 2003, Jeremy Hansen wrote:
> > I'm in the process of planning a gigE network and I'd like to get the
> > latest on performance tuning for gigE copper and what cards to look at for
> > maximum performance. If there is a gigE FAQ, please point me to it.
> No good FAQ that I know of. It's much more difficult to get information
> about the internal design of GbE adapters than it was a few years ago.
> > Right now what I have available to play with is a Cisco 6509 with a 48port
> > 10/100/1000 copper blade and a few e1000's.
> The exact model of the e1000 is important. The rapid churn of PCI ID
> numbers for the e1000 cards reflects the changes in the internal
> structure and firmware of the chips. The e1000 cards are much smarter
> about looking at the payload of packets and basing interrupt decisions
> on the contents. This is good for latency and typical performance, but
> makes it difficult to do micro-benchmarks.
03:0e.0 Ethernet controller: Intel Corp. 82545EM Gigabit Ethernet
Controller (Copper) (rev 01)
01:02.0 Ethernet controller: Intel Corp. 82544EI Gigabit Ethernet
Controller (rev 02)
> > I'd like to benchmark the
> > different speeds I'm getting while playing with MTU's on these machines
> > but netperf doesn't seem to like it when I set to MTU to 9000. Ttcp seems
> > to do the same. Our goal is to see somewhere around 500M/s of raw TCP
> > between our servers and workstations.
> Getting jumbo frames to work is tricky. Jumbo frames are not part of
> the standard, for good technical reasons.
The Catalyst 6509 we have seems to support this. It was only 'til a few
moment ago that I realized that in order to do jumbo frames, that both the
machine and the switch need to support this.
> Have you configured the switch to match the exact size configured on the
> Linux machine?
This is the next step. I'll report back when I have jumbo frames enable
on both the switch and the server.