On Thursday 06 January 2005 02:29 pm, Robert Olsson wrote:
> Jeremy M. Guthrie writes:
> > > You only use CPU0 for packet processing. Also it seems you use the
> > > non-NAPI version of e1000.
> >
> > The E1000 driver is the stock driver in 2.4.28.
> There is a kernel config option. Use Rx Polling (NAPI)
I'm recompiling the module and will reload the module.
> > eth2 is TX only. We don't receive anything on it. This system should
> > only ever RX on eth3 and TX on eth2 as part of its function.
>
> Ok! So unidirectional traffic and CPU0 processes all skb's and passes them
> over to CPU1 which touches and frees them. You could try some other
> affinity setting later on but there are so many options depending what
> we want do. If we just want to compare 2.4/2.6 we can start out simple
> with setting affinity so only one CPU. Which is almost what you got.
> We reach all the complicated problems immediately... Now we pass skb's
> between the CPU's which release cache bouncing and makes slab rebalance
> it's pools. So using just one CPU is probably better... If we had incoming
> pkts on eth2 eventually we could have had any use for CPU1.
>
> Install NAPI-driver, set affinity so eth2/eth3 uses same CPU to start
> with.
I installed the NAPI driver and now I drop packets under 2.4.28. I tried
setting eth2/3 onto the same CPU. If both eth2/3 are on CPU0, I drop a lot
of packets. If CPU0-eth2 & CPU1-eth3, I still drop packets, just not as
fast.
I am going to try V2.6 w/o NAPI to see how much that helps.
> The stats info you sent was almost impossible to read. See if you can get
> only the rtstats.
--
--------------------------------------------------
Jeremy M. Guthrie jeremy.guthrie@xxxxxxxxxx
Senior Network Engineer Phone: 608-298-1061
Berbee Fax: 608-288-3007
5520 Research Park Drive NOC: 608-298-1102
Madison, WI 53711
pgpTZvyttfOHw.pgp
Description: PGP signature
|