Hi all,
I found that testing if any tx work is done in dev->poll before exiting
polled mode improves performance by about 7% (max) in a 2-port e100 bridge
forwarding unidirectional test case. If tx work is not considered when
deciding whether to netif_rx_complete, the transmitting interface sees
loads of interrupts and hence forwarding throughput is degraded.
When testing with bidirectional test data, no improvement is seen since
the dev->poll is kept in polled mode on both interfaces due to
receive work.
Hope this helps.
-jc
> > > Furthermore NAPI can be extended to schedule dev->poll even for TX-
> > > interrupts. There is pacth for e1000 doing this. We see about 5-8%
> > > overall system packet improvement with this.
> > I was thinking of a variant JC [jchapman@xxxxxxxxxxx]
> > mentioned on this list some time ago. He also sent me
> > the patch for e1000. A test and the patch is below.
>
> JC contributed almost the exact patch for the e100 rewrite and it did
> help Tx, but I don't remember how much. JC, do you remember? Here is
> the snippet:
>
> static int e100_poll(struct net_device *netdev, int *budget)
> {
> struct nic *nic = netdev->priv;
> unsigned int work_to_do = min(netdev->quota, *budget);
> unsigned int work_done = 0;
> int tx_cleaned;
>
> e100_rx_clean(nic, &work_done, work_to_do);
> tx_cleaned = e100_tx_clean(nic);
>
> /* If no Rx and Tx cleanup work was done, exit polling mode. */
> if((!tx_cleaned && (work_done == 0)) || !netif_running(netdev))
> {
> netif_rx_complete(netdev);
> e100_enable_irq(nic);
> return 0;
> }
>
> *budget -= work_done;
> netdev->quota -= work_done;
>
> return 1;
> }
-------------------------------------------------
This mail sent through IMP: http://horde.org/imp/
|