Stephen Hemminger writes:
> > On 11/20/05, Unai Uribarri <unaiur@xxxxxxxxx> wrote:
> > > I'm trying to run a gigabit linux-based bridge at full duplex, full
> > > line rate. That means receiving and sending 3 Million Packets Per
> > > Second: quite ambitious, I known.
> >
> > 3.125Mpps, actually, assuming all minimum length 40 byte packets.
> >
> > > My first problem is to get the appropiate hardware: I've evaluated
> > > several PCI-X nics (Intel & Broadcom) but I can't achieve more than 800
> > > kpps. I've read that this is the PCI-X bus limit, so I'm going to buy a
> > > pair of PCI Express x4 NICs.
Yes PCI-X boards is sending around ~800 kpps. With some ugly patches about
300 kpps more.
Intel w. e1000 82546GB @ 133 MHz
60 748406
124 693133
252 452951
508 234982
1020 119742
1496 82242
The BCM PCI-X *inside* the serverworks H2000 is faster.
BCM w. tg3
60 1421959
124 844572
252 452918
508 234970
1020 119735
1496 82239
Fastest and most intresing sofar is the Intel 82571 Server PCI-E dual
server adapter.
60 1488305
124 844635
252 452835
508 234973
1020 119737
1496 82240
IO latency is worse with PCI-E but the Intel 82571 can handle four
concurrent RX/TX transactions. Also PCI-E is FX this seems to well
handle the PCI-E extra latency.
Stephen I've heard some good things about the Syskonnect PCI-E adapters
any chance you could run a test similar to the tests above?
> Robert Olsson had a patch to count packets sent to discard port in the
> receive path for for measuring packet rate.
Yes if are just to drop pktgen pkts in e1000 driver. You can use the
patch below. It can be done a little bit better. See skb reuse
discussion on netdev. In that case we reuse the skb in e1000 driver
as well and saves kmalloc/kfree's.
Cheers.
--ro
--- drivers/net/e1000/e1000_main.c~ 2003-10-16 11:25:23.000000000 +0200
+++ drivers/net/e1000/e1000_main.c 2003-10-16 13:24:46.000000000 +0200
@@ -2609,6 +2609,23 @@
continue;
}
}
+#define DMA_TEST
+#ifdef DMA_TEST
+ {
+ __u8 *data = (__u8 *) skb->data + 34; /* 14+20 */
+
+ /* src and dst port 9 --> pktgen */
+
+ if(data[0] == 0 &&
+ data[1] == 9 &&
+ data[2] == 0 &&
+ data[3] == 9) {
+ dev_kfree_skb(skb);
+ adapter->net_stats.rx_compressed++;
+ goto gone;
+ }
+ }
+#endif
/* Good Receive */
skb_put(skb, length - ETHERNET_FCS_SIZE);
@@ -2634,6 +2651,10 @@
netif_rx(skb);
}
#endif /* CONFIG_E1000_NAPI */
+
+#ifdef DMA_TEST
+gone:
+#endif
netdev->last_rx = jiffies;
rx_desc->status = 0;
-
: send the line "unsubscribe linux-net" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
